mp-juuuns commited on
Commit
1bc5bc2
·
verified ·
1 Parent(s): 7121fa8

Put the full metric comparison at the top; strip local paths; drop stale manifests

Browse files

Comparison: the card never placed this root against the field. It now does, at the top, on every metric the protocol measured. Quality rank 3 of 65 (a six-arm tie inside one median seed SD), latency 52/65, energy 52/65, RSS 40/65, file 31/65. New figure 10 shows all 65 arms per metric.

Local paths: seven manifests carried /home/mp_ubuntu/Projects/... absolute paths into a public repository. 24 occurrences rewritten to repo-relative; no other content changed.

Deleted, all stale or wrong:
release_manifest.json - asserts the OLD repo id and the OLD root's weight sha256
benchmark/summary.json - superseded-lineage numbers; also claims a test split this repo
benchmark/semeval-transfer-summary.json does not have. Byte-identical triplicate.
benchmark/figures/source_summary.json
benchmark/figures/manifest.json - schema standalone4l-v128k-figures-v1, describes a
benchmark/figures/cross_task/manifest.json figure set the current card does not use
benchmark/figures/cross_task/source_summary.json

.gitattributes CHANGED
@@ -11,3 +11,4 @@ benchmark/figures/07_runtime_specialization.png filter=lfs diff=lfs merge=lfs -t
11
  benchmark/figures/08_long_context_window.png filter=lfs diff=lfs merge=lfs -text
12
  docker/bundle/model/classifier-q8_0.gguf filter=lfs diff=lfs merge=lfs -text
13
  docker/bundle/runtime_source.tar.gz filter=lfs diff=lfs merge=lfs -text
 
 
11
  benchmark/figures/08_long_context_window.png filter=lfs diff=lfs merge=lfs -text
12
  docker/bundle/model/classifier-q8_0.gguf filter=lfs diff=lfs merge=lfs -text
13
  docker/bundle/runtime_source.tar.gz filter=lfs diff=lfs merge=lfs -text
14
+ benchmark/figures/10_root_on_every_metric.png filter=lfs diff=lfs merge=lfs -text
benchmark/figures/10_root_on_every_metric.png ADDED

Git LFS Details

  • SHA256: 56a7353a71dd727077a420fd39a97b21742b8af87fabcbd4d4cdbfe09f5a0b40
  • Pointer size: 131 Bytes
  • Size of remote file: 158 kB
benchmark/figures/cross_task/manifest.json DELETED
@@ -1,158 +0,0 @@
1
- {
2
- "figures": [
3
- "05_why_4l_edge_tradeoff",
4
- "06_malint_4l_quality",
5
- "07_malint_quality_resource"
6
- ],
7
- "files_before_manifest": [
8
- {
9
- "bytes": 5442,
10
- "path": "05_why_4l_edge_tradeoff.csv",
11
- "sha256": "afd203646513c49140a25ef2a562a064286a4fa5efb495e6a73f83f0106e7933"
12
- },
13
- {
14
- "bytes": 27224,
15
- "path": "05_why_4l_edge_tradeoff.pdf",
16
- "sha256": "22f531ea586f9c79ac8f114345cec56e74ea5ed97aa3726f7310018666ae6312"
17
- },
18
- {
19
- "bytes": 178965,
20
- "path": "05_why_4l_edge_tradeoff.png",
21
- "sha256": "cdc0a8a98cd13ad35a68b980d2e0750e2424b8932f7d27dc0ecd548fed823c74"
22
- },
23
- {
24
- "bytes": 2118,
25
- "path": "06_malint_4l_quality.csv",
26
- "sha256": "25848237b9c77596a8c4b6224fa111f86c816e065804ccfdd6a4fdf6c63ef8d1"
27
- },
28
- {
29
- "bytes": 26351,
30
- "path": "06_malint_4l_quality.pdf",
31
- "sha256": "9a1ae78b4e5871f5d5c5321db6c5d1fe0bf349a7869977b2649b44ccea633d04"
32
- },
33
- {
34
- "bytes": 181392,
35
- "path": "06_malint_4l_quality.png",
36
- "sha256": "2c5b22415d311ca5fa2b198d55b77f3570b57595d01f0b5031abde59ffc3e9e6"
37
- },
38
- {
39
- "bytes": 1800,
40
- "path": "07_malint_quality_resource.csv",
41
- "sha256": "aac7d3400cf8dcb3ac8b6fc317c27ded5343231c93476997011a3bb56b5e4d6c"
42
- },
43
- {
44
- "bytes": 29309,
45
- "path": "07_malint_quality_resource.pdf",
46
- "sha256": "498338991976e64038096a12361f63830ced3b9c6eea3e1bdaed8d0e4efac606"
47
- },
48
- {
49
- "bytes": 163562,
50
- "path": "07_malint_quality_resource.png",
51
- "sha256": "1cb13cc695f3d59ae9ce75599b86867da624d6e0aae0897e0db8686f6db984ba"
52
- },
53
- {
54
- "bytes": 868,
55
- "path": "README.md",
56
- "sha256": "f3e79965f14ec0fd20d11c361d25c3a3abc4c9a20b5dbce9fac0660bb12789f0"
57
- },
58
- {
59
- "bytes": 29752,
60
- "path": "source_summary.json",
61
- "sha256": "20ea292f85757440d8641199003c47674deb0c6aacea0ef8f9f5fcd2480bb5d0"
62
- }
63
- ],
64
- "inputs": [
65
- {
66
- "bytes": 58857,
67
- "path": "results/malint_qwen35_4l_repeated_cv/20260814-r3x3-terminal-v1/aggregate-v1/aggregate_result.json",
68
- "role": "malint_repeated_aggregate",
69
- "sha256": "e3d659c2f5aca645f040d03c57e4effa9eb2616f622aceb503c71236b6b01efa"
70
- },
71
- {
72
- "bytes": 325527,
73
- "path": "results/malint_qwen35_4l_repeated_cv/20260814-r3x3-terminal-v1/aggregate-v1/aggregate_oof_predictions.npz",
74
- "role": "malint_repeated_oof",
75
- "sha256": "a1d7827ebe4beecb49676ca42e0c986c5f4101c394d55c9965161481c347a3f1"
76
- },
77
- {
78
- "bytes": 15407,
79
- "path": "results/malint_qwen35_4l_resource_benchmark/20260814-matched-bf16-v2/resource_benchmark.json",
80
- "role": "malint_resource_aggregate",
81
- "sha256": "5337e44c907a05590c0ea3da6a0760baff22c3f930ed81c069c257060fd2e725"
82
- },
83
- {
84
- "bytes": 3144,
85
- "path": "results/final_broad_benchmark/20260729-v1/report/quality_matrix.csv",
86
- "role": "propaganda_quality_matrix",
87
- "sha256": "c6e316f7081e2a6dcda1ec4e95b57cdcc357ae49ed8b3f5a876c5b51e0df722f"
88
- },
89
- {
90
- "bytes": 115661,
91
- "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-full24l-commonhead/seed41/report.json",
92
- "role": "propaganda_seed_report_01",
93
- "sha256": "1c572fb728755dbde0460e6b8eb9a84d078b128f8a42a9148a17c40fc0b27d82"
94
- },
95
- {
96
- "bytes": 116024,
97
- "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-full24l-commonhead/seed42/report.json",
98
- "role": "propaganda_seed_report_02",
99
- "sha256": "055edfee92e39928aa9577f239c5b0389f3dda375a96fc69a2c000a0ed6edc36"
100
- },
101
- {
102
- "bytes": 115639,
103
- "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-full24l-commonhead/seed43-retry2/report.json",
104
- "role": "propaganda_seed_report_03",
105
- "sha256": "84b2c316556db4aa5fdf7c08433a4808a5fe9ef6f620da6472f4f2f73dee2e4b"
106
- },
107
- {
108
- "bytes": 114924,
109
- "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-c8q8l-commonhead/seed41/report.json",
110
- "role": "propaganda_seed_report_04",
111
- "sha256": "9234258589ffb24eaa92859639a66aeb3e55190f8aa82f798d88e791a282bf9c"
112
- },
113
- {
114
- "bytes": 115800,
115
- "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-c8q8l-commonhead/seed42/report.json",
116
- "role": "propaganda_seed_report_05",
117
- "sha256": "8161dd1378081de9960136fddab800994529aa3c754674ca41bb4aa71515b3c6"
118
- },
119
- {
120
- "bytes": 116039,
121
- "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-c8q8l-commonhead/seed43/report.json",
122
- "role": "propaganda_seed_report_06",
123
- "sha256": "ce0c76a2aa97a5b7c8cdf820af0bd4f8a26bf50d460612ca1db30bd497ca3d9d"
124
- },
125
- {
126
- "bytes": 115985,
127
- "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-standalone4l-commonhead/seed41/report.json",
128
- "role": "propaganda_seed_report_07",
129
- "sha256": "ec5314140fff0bdd46e87ee00ad67bee7e61d82c7cfa34add1510dd7d5bdb3c0"
130
- },
131
- {
132
- "bytes": 115972,
133
- "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-standalone4l-commonhead/seed42/report.json",
134
- "role": "propaganda_seed_report_08",
135
- "sha256": "e5c45a7dc41bebf4550da7c5763d0759c6d0922da9efda4048564c11c6d2f3af"
136
- },
137
- {
138
- "bytes": 116177,
139
- "path": "results/final_broad_benchmark/20260729-v1/trainable/qwen35-standalone4l-commonhead/seed43/report.json",
140
- "role": "propaganda_seed_report_09",
141
- "sha256": "e05bc2146d720706780fc45a1fb05911d5c2926caa0bcb12ffead946028eef0d"
142
- },
143
- {
144
- "bytes": 827,
145
- "path": "results/standalone4l_base/20260731-v1/huggingface_release_v3/benchmark/compression_ladder.csv",
146
- "role": "task_free_compression_ladder",
147
- "sha256": "2c544c3907c4f592d724f658850b8c6b9634845cb5b50a66ebd2ea6c5d48ce0b"
148
- },
149
- {
150
- "bytes": 13963,
151
- "path": "results/unoq_lineage_matrix/20260731-v1/analysis/analysis.json",
152
- "role": "uno_q_depth_analysis",
153
- "sha256": "c899a19e4433e3a40547a11cfcfc3b0a7b6f00c436aa8c1d742a5dc0399bf649"
154
- }
155
- ],
156
- "schema": "qwen35-4l-cross-task-base-card-figures-manifest-v1",
157
- "status": "complete"
158
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
benchmark/figures/cross_task/source_summary.json DELETED
@@ -1,972 +0,0 @@
1
- {
2
- "claim_boundaries": [
3
- "The controlled UNO Q 4L point is a historical early-path proxy, not the final common-head quality endpoint.",
4
- "The propaganda 24L/8L/4L endpoints differ in lineage and vocabulary; their comparison is descriptive, not a pure depth ablation.",
5
- "All Qwen MALINT student arms are 4L; MALINT supports a 4L interface result, not 4L-versus-8L/24L depth superiority.",
6
- "The MALINT result is post-selection internal repeated-fold evidence, not untouched held-out confirmation or SOTA evidence.",
7
- "Mean+KD incurred measured CUDA-forward latency and peak-allocation increases; it is not claimed to be faster.",
8
- "No ambient GPU utilization value is used as model occupancy."
9
- ],
10
- "malint_quality": [
11
- {
12
- "arm": "rightmost_hard",
13
- "arm_label": "Rightmost + hard",
14
- "layers": 4,
15
- "mean": 0.522182806412052,
16
- "metric": "weighted_f1",
17
- "n_repeats": 3,
18
- "role": "qwen_4l_student",
19
- "sample_sd": 0.014974460565445307,
20
- "threshold": 0.3
21
- },
22
- {
23
- "arm": "rightmost_hard",
24
- "arm_label": "Rightmost + hard",
25
- "layers": 4,
26
- "mean": 0.5109624469249431,
27
- "metric": "macro_f1",
28
- "n_repeats": 3,
29
- "role": "qwen_4l_student",
30
- "sample_sd": 0.016275524962473233,
31
- "threshold": 0.3
32
- },
33
- {
34
- "arm": "rightmost_hard",
35
- "arm_label": "Rightmost + hard",
36
- "layers": 4,
37
- "mean": 0.5320256865023298,
38
- "metric": "micro_f1",
39
- "n_repeats": 3,
40
- "role": "qwen_4l_student",
41
- "sample_sd": 0.010757300158510568,
42
- "threshold": 0.3
43
- },
44
- {
45
- "arm": "rightmost_hard",
46
- "arm_label": "Rightmost + hard",
47
- "layers": 4,
48
- "mean": 0.6403030303030303,
49
- "metric": "subset_exact_match",
50
- "n_repeats": 3,
51
- "role": "qwen_4l_student",
52
- "sample_sd": 0.003674047167570365,
53
- "threshold": 0.3
54
- },
55
- {
56
- "arm": "rightmost_hard",
57
- "arm_label": "Rightmost + hard",
58
- "layers": 4,
59
- "mean": 0.8746666666666667,
60
- "metric": "label_accuracy",
61
- "n_repeats": 3,
62
- "role": "qwen_4l_student",
63
- "sample_sd": 0.0012106051124507965,
64
- "threshold": 0.3
65
- },
66
- {
67
- "arm": "mean_hard",
68
- "arm_label": "Mean pool + hard",
69
- "layers": 4,
70
- "mean": 0.6118049035859879,
71
- "metric": "weighted_f1",
72
- "n_repeats": 3,
73
- "role": "qwen_4l_student",
74
- "sample_sd": 0.008860176348616246,
75
- "threshold": 0.3
76
- },
77
- {
78
- "arm": "mean_hard",
79
- "arm_label": "Mean pool + hard",
80
- "layers": 4,
81
- "mean": 0.6055064323204021,
82
- "metric": "macro_f1",
83
- "n_repeats": 3,
84
- "role": "qwen_4l_student",
85
- "sample_sd": 0.007533985719586246,
86
- "threshold": 0.3
87
- },
88
- {
89
- "arm": "mean_hard",
90
- "arm_label": "Mean pool + hard",
91
- "layers": 4,
92
- "mean": 0.6176900341883843,
93
- "metric": "micro_f1",
94
- "n_repeats": 3,
95
- "role": "qwen_4l_student",
96
- "sample_sd": 0.007461802499370809,
97
- "threshold": 0.3
98
- },
99
- {
100
- "arm": "mean_hard",
101
- "arm_label": "Mean pool + hard",
102
- "layers": 4,
103
- "mean": 0.6696969696969697,
104
- "metric": "subset_exact_match",
105
- "n_repeats": 3,
106
- "role": "qwen_4l_student",
107
- "sample_sd": 0.005169309730070291,
108
- "threshold": 0.3
109
- },
110
- {
111
- "arm": "mean_hard",
112
- "arm_label": "Mean pool + hard",
113
- "layers": 4,
114
- "mean": 0.8876363636363637,
115
- "metric": "label_accuracy",
116
- "n_repeats": 3,
117
- "role": "qwen_4l_student",
118
- "sample_sd": 0.0017533910474532954,
119
- "threshold": 0.3
120
- },
121
- {
122
- "arm": "mean_kd",
123
- "arm_label": "Mean pool + KD",
124
- "layers": 4,
125
- "mean": 0.6346533875063399,
126
- "metric": "weighted_f1",
127
- "n_repeats": 3,
128
- "role": "qwen_4l_student",
129
- "sample_sd": 0.011966996735426847,
130
- "threshold": 0.3
131
- },
132
- {
133
- "arm": "mean_kd",
134
- "arm_label": "Mean pool + KD",
135
- "layers": 4,
136
- "mean": 0.6281768437686408,
137
- "metric": "macro_f1",
138
- "n_repeats": 3,
139
- "role": "qwen_4l_student",
140
- "sample_sd": 0.014225316120600004,
141
- "threshold": 0.3
142
- },
143
- {
144
- "arm": "mean_kd",
145
- "arm_label": "Mean pool + KD",
146
- "layers": 4,
147
- "mean": 0.6389632888388834,
148
- "metric": "micro_f1",
149
- "n_repeats": 3,
150
- "role": "qwen_4l_student",
151
- "sample_sd": 0.009850482866928375,
152
- "threshold": 0.3
153
- },
154
- {
155
- "arm": "mean_kd",
156
- "arm_label": "Mean pool + KD",
157
- "layers": 4,
158
- "mean": 0.6557575757575758,
159
- "metric": "subset_exact_match",
160
- "n_repeats": 3,
161
- "role": "qwen_4l_student",
162
- "sample_sd": 0.003192622349349309,
163
- "threshold": 0.3
164
- },
165
- {
166
- "arm": "mean_kd",
167
- "arm_label": "Mean pool + KD",
168
- "layers": 4,
169
- "mean": 0.8757575757575757,
170
- "metric": "label_accuracy",
171
- "n_repeats": 3,
172
- "role": "qwen_4l_student",
173
- "sample_sd": 0.0020436018545560413,
174
- "threshold": 0.3
175
- },
176
- {
177
- "arm": "teacher",
178
- "arm_label": "RoBERTa teacher",
179
- "layers": "",
180
- "mean": 0.6046800630077942,
181
- "metric": "weighted_f1",
182
- "n_repeats": 3,
183
- "role": "external_teacher",
184
- "sample_sd": 0.01579900095351724,
185
- "threshold": 0.3
186
- },
187
- {
188
- "arm": "teacher",
189
- "arm_label": "RoBERTa teacher",
190
- "layers": "",
191
- "mean": 0.5955721586042237,
192
- "metric": "macro_f1",
193
- "n_repeats": 3,
194
- "role": "external_teacher",
195
- "sample_sd": 0.01637303032583919,
196
- "threshold": 0.3
197
- },
198
- {
199
- "arm": "teacher",
200
- "arm_label": "RoBERTa teacher",
201
- "layers": "",
202
- "mean": 0.6046188233688233,
203
- "metric": "micro_f1",
204
- "n_repeats": 3,
205
- "role": "external_teacher",
206
- "sample_sd": 0.01776761734134435,
207
- "threshold": 0.3
208
- },
209
- {
210
- "arm": "teacher",
211
- "arm_label": "RoBERTa teacher",
212
- "layers": "",
213
- "mean": 0.62,
214
- "metric": "subset_exact_match",
215
- "n_repeats": 3,
216
- "role": "external_teacher",
217
- "sample_sd": 0.005961307749365459,
218
- "threshold": 0.3
219
- },
220
- {
221
- "arm": "teacher",
222
- "arm_label": "RoBERTa teacher",
223
- "layers": "",
224
- "mean": 0.8507878787878788,
225
- "metric": "label_accuracy",
226
- "n_repeats": 3,
227
- "role": "external_teacher",
228
- "sample_sd": 0.01087469324477906,
229
- "threshold": 0.3
230
- }
231
- ],
232
- "malint_quality_contract": {
233
- "claim_boundary": "Post-selection internal repeated-fold robustness and matched ablation only. This result is not untouched confirmation, official held-out evidence, all-label superiority, or SOTA evidence.",
234
- "contrasts": {
235
- "conditional_kd_mean_kd_minus_mean_hard": {
236
- "corrected_resampled_t_interval": {
237
- "confidence": 0.95,
238
- "critical_t": 2.306004135204166,
239
- "degrees_of_freedom": 8,
240
- "formula": "sqrt((1/n + mean(n_eval/n_train)) * sample_variance)",
241
- "interval": [
242
- -0.0015478418507784523,
243
- 0.05059146195651548
244
- ],
245
- "mean": 0.024521810052868513,
246
- "mean_eval_train_ratio": 0.5000006195533516,
247
- "n": 9,
248
- "sample_variance": 0.00020913642539219796,
249
- "standard_error": 0.011305119321192738,
250
- "variance_correction": 0.6111117306644627
251
- },
252
- "fold_deltas": [
253
- 0.04524702943141845,
254
- 0.017956307877355804,
255
- -0.0053068243793217285,
256
- 0.015864642448092825,
257
- 0.022064206421293142,
258
- 0.03213041489613799,
259
- 0.02962871459506977,
260
- 0.026719118927120977,
261
- 0.03639268025864939
262
- ],
263
- "fold_mean": 0.024521810052868513,
264
- "fold_sample_standard_deviation": 0.014461549895920491,
265
- "name": "conditional_kd_mean_kd_minus_mean_hard",
266
- "per_label_mean_delta": {
267
- "CPV": 0.0006438268748058898,
268
- "PASV": 0.027106289374744647,
269
- "PSSA": 0.058741661936162225,
270
- "UCPI": 0.0247139382339762,
271
- "UIOA": 0.0021463408215049995
272
- },
273
- "per_label_repeat_deltas": [
274
- [
275
- -0.01842723004694835,
276
- 0.04865531704909942,
277
- 0.015803620583568212,
278
- -0.013921424268790505,
279
- 0.03679497238444562
280
- ],
281
- [
282
- 0.025183733463988223,
283
- 0.070038654494613,
284
- -0.0026296349009481412,
285
- 0.021790722761596548,
286
- 0.0004450534649754756
287
- ],
288
- [
289
- -0.004825022792622202,
290
- 0.05753101426477425,
291
- -0.006734963218105072,
292
- 0.0734495696314279,
293
- 0.0369017888525075
294
- ]
295
- ],
296
- "positive_fold_tail": {
297
- "null_probability": 0.5,
298
- "one_sided_probability_at_least_observed": 0.01953125,
299
- "positive": 8,
300
- "total": 9
301
- },
302
- "repeat_oof_deltas": [
303
- 0.018545164631776245,
304
- 0.020840853227636957,
305
- 0.029159433901642884
306
- ],
307
- "repeat_oof_t_interval": {
308
- "confidence": 0.95,
309
- "critical_t": 4.302652729749462,
310
- "degrees_of_freedom": 2,
311
- "interval": [
312
- 0.00897537793561895,
313
- 0.03672158990508511
314
- ],
315
- "mean": 0.02284848392035203,
316
- "n": 3,
317
- "sample_standard_deviation": 0.005584676694496962,
318
- "standard_error": 0.0032243145929048505
319
- }
320
- },
321
- "mean_kd_minus_teacher": {
322
- "corrected_resampled_t_interval": {
323
- "confidence": 0.95,
324
- "critical_t": 2.306004135204166,
325
- "degrees_of_freedom": 8,
326
- "formula": "sqrt((1/n + mean(n_eval/n_train)) * sample_variance)",
327
- "interval": [
328
- -0.03180974553628697,
329
- 0.08853586985785872
330
- ],
331
- "mean": 0.02836306216078588,
332
- "mean_eval_train_ratio": 0.5000006195533516,
333
- "n": 9,
334
- "sample_variance": 0.0011141913151793519,
335
- "standard_error": 0.026093972156622066,
336
- "variance_correction": 0.6111117306644627
337
- },
338
- "fold_deltas": [
339
- -0.03162698747672954,
340
- 0.06379890775996899,
341
- 0.056964918230746675,
342
- 0.04693837159160652,
343
- 0.001345287770746495,
344
- 0.014058020214096523,
345
- 0.00030085368163068615,
346
- 0.05808798200431842,
347
- 0.04540020567068814
348
- ],
349
- "fold_mean": 0.02836306216078588,
350
- "fold_sample_standard_deviation": 0.0333795044178213,
351
- "name": "mean_kd_minus_teacher",
352
- "per_label_mean_delta": {
353
- "CPV": 0.03245133981428311,
354
- "PASV": 0.04215462872506418,
355
- "PSSA": 0.030550793916265178,
356
- "UCPI": 0.006157663285133859,
357
- "UIOA": 0.05170900008133931
358
- },
359
- "per_label_repeat_deltas": [
360
- [
361
- 0.010081541882876188,
362
- 0.09069144184384492,
363
- 0.058604490711147705,
364
- -0.005178527872112815,
365
- 0.0012970008399624255
366
- ],
367
- [
368
- 0.05816392692322364,
369
- -0.013100009776126686,
370
- 0.017979371387083054,
371
- 0.03578643578643581,
372
- 0.01576007056748019
373
- ],
374
- [
375
- 0.029108550636749497,
376
- 0.014060949681077295,
377
- 0.07854313814578717,
378
- 0.09585597826086956,
379
- 0.0014159184479589637
380
- ]
381
- ],
382
- "positive_fold_tail": {
383
- "null_probability": 0.5,
384
- "one_sided_probability_at_least_observed": 0.01953125,
385
- "positive": 8,
386
- "total": 9
387
- },
388
- "repeat_oof_deltas": [
389
- 0.032162246558158025,
390
- 0.020416569493259984,
391
- 0.037341157444219175
392
- ],
393
- "repeat_oof_t_interval": {
394
- "confidence": 0.95,
395
- "critical_t": 4.302652729749462,
396
- "degrees_of_freedom": 2,
397
- "interval": [
398
- 0.00843082924454409,
399
- 0.05151581975254736
400
- ],
401
- "mean": 0.029973324498545728,
402
- "n": 3,
403
- "sample_standard_deviation": 0.008672021342497449,
404
- "standard_error": 0.005006793856509082
405
- }
406
- },
407
- "pooling_mean_hard_minus_rightmost_hard": {
408
- "corrected_resampled_t_interval": {
409
- "confidence": 0.95,
410
- "critical_t": 2.306004135204166,
411
- "degrees_of_freedom": 8,
412
- "formula": "sqrt((1/n + mean(n_eval/n_train)) * sample_variance)",
413
- "interval": [
414
- 0.056762831891190316,
415
- 0.12504455147590826
416
- ],
417
- "mean": 0.0909036916835493,
418
- "mean_eval_train_ratio": 0.5000006195533516,
419
- "n": 9,
420
- "sample_variance": 0.00035868079546522353,
421
- "standard_error": 0.014805203195932804,
422
- "variance_correction": 0.6111117306644627
423
- },
424
- "fold_deltas": [
425
- 0.06678423059606675,
426
- 0.08178035353463753,
427
- 0.08672762777739151,
428
- 0.09344774812516743,
429
- 0.0878998694863643,
430
- 0.07415570161139795,
431
- 0.11094545718958193,
432
- 0.08708564234043492,
433
- 0.12930659449090132
434
- ],
435
- "fold_mean": 0.0909036916835493,
436
- "fold_sample_standard_deviation": 0.01893886996273071,
437
- "name": "pooling_mean_hard_minus_rightmost_hard",
438
- "per_label_mean_delta": {
439
- "CPV": 0.14860808426516173,
440
- "PASV": 0.09778905085334237,
441
- "PSSA": 0.07923875253833933,
442
- "UCPI": 0.055020095369494,
443
- "UIOA": 0.09206394395095778
444
- },
445
- "per_label_repeat_deltas": [
446
- [
447
- 0.09436936936936935,
448
- 0.07977821265492502,
449
- 0.10417829690385194,
450
- 0.07872402692420255,
451
- 0.03863953001237319
452
- ],
453
- [
454
- 0.15668016194331985,
455
- 0.04167051934042232,
456
- 0.06934192200557099,
457
- 0.11942791523123497,
458
- 0.06908520415955133
459
- ],
460
- [
461
- 0.194774721482796,
462
- 0.11626752561967063,
463
- 0.1026716129434504,
464
- 0.09521521040458958,
465
- 0.05733555193655748
466
- ]
467
- ],
468
- "positive_fold_tail": {
469
- "null_probability": 0.5,
470
- "one_sided_probability_at_least_observed": 0.001953125,
471
- "positive": 9,
472
- "total": 9
473
- },
474
- "repeat_oof_deltas": [
475
- 0.07574451062131737,
476
- 0.08540665760778032,
477
- 0.10771512329271016
478
- ],
479
- "repeat_oof_t_interval": {
480
- "confidence": 0.95,
481
- "critical_t": 4.302652729749462,
482
- "degrees_of_freedom": 2,
483
- "interval": [
484
- 0.048890004893968025,
485
- 0.13035418945390387
486
- ],
487
- "mean": 0.08962209717393595,
488
- "n": 3,
489
- "sample_standard_deviation": 0.01639687368671199,
490
- "standard_error": 0.009466739436891459
491
- }
492
- },
493
- "primary_mean_kd_minus_rightmost_hard": {
494
- "corrected_resampled_t_interval": {
495
- "confidence": 0.95,
496
- "critical_t": 2.306004135204166,
497
- "degrees_of_freedom": 8,
498
- "formula": "sqrt((1/n + mean(n_eval/n_train)) * sample_variance)",
499
- "interval": [
500
- 0.07160480852626178,
501
- 0.15924619494657383
502
- ],
503
- "mean": 0.1154255017364178,
504
- "mean_eval_train_ratio": 0.5000006195533516,
505
- "n": 9,
506
- "sample_variance": 0.0005909050521203029,
507
- "standard_error": 0.019002868442938105,
508
- "variance_correction": 0.6111117306644627
509
- },
510
- "fold_deltas": [
511
- 0.1120312600274852,
512
- 0.09973666141199333,
513
- 0.08142080339806979,
514
- 0.10931239057326025,
515
- 0.10996407590765744,
516
- 0.10628611650753594,
517
- 0.1405741717846517,
518
- 0.1138047612675559,
519
- 0.1656992747495507
520
- ],
521
- "fold_mean": 0.1154255017364178,
522
- "fold_sample_standard_deviation": 0.024308538666902683,
523
- "name": "primary_mean_kd_minus_rightmost_hard",
524
- "per_label_mean_delta": {
525
- "CPV": 0.14925191113996764,
526
- "PASV": 0.12489534022808702,
527
- "PSSA": 0.13798041447450154,
528
- "UCPI": 0.0797340336034702,
529
- "UIOA": 0.09421028477246278
530
- },
531
- "per_label_repeat_deltas": [
532
- [
533
- 0.075942139322421,
534
- 0.12843352970402444,
535
- 0.11998191748742015,
536
- 0.06480260265541204,
537
- 0.07543450239681881
538
- ],
539
- [
540
- 0.18186389540730807,
541
- 0.11170917383503531,
542
- 0.06671228710462285,
543
- 0.14121863799283152,
544
- 0.0695302576245268
545
- ],
546
- [
547
- 0.1899496986901738,
548
- 0.17379853988444488,
549
- 0.09593664972534532,
550
- 0.16866478003601748,
551
- 0.09423734078906498
552
- ]
553
- ],
554
- "positive_fold_tail": {
555
- "null_probability": 0.5,
556
- "one_sided_probability_at_least_observed": 0.001953125,
557
- "positive": 9,
558
- "total": 9
559
- },
560
- "repeat_oof_deltas": [
561
- 0.09428967525309362,
562
- 0.10624751083541728,
563
- 0.13687455719435304
564
- ],
565
- "repeat_oof_t_interval": {
566
- "confidence": 0.95,
567
- "critical_t": 4.302652729749462,
568
- "degrees_of_freedom": 2,
569
- "interval": [
570
- 0.05790922361228608,
571
- 0.1670319385762899
572
- ],
573
- "mean": 0.11247058109428798,
574
- "n": 3,
575
- "sample_standard_deviation": 0.02196390209122417,
576
- "standard_error": 0.01268086478482286
577
- }
578
- }
579
- },
580
- "promotion_gate": {
581
- "fold_gate_passed": true,
582
- "observed_positive_fold_deltas": 9,
583
- "observed_positive_repeat_oof_deltas": 3,
584
- "passed": true,
585
- "repeat_gate_passed": true,
586
- "required_positive_fold_deltas": 7,
587
- "required_positive_repeat_oof_deltas": 3
588
- },
589
- "study": {
590
- "fold_count": 9,
591
- "fold_runner_protocol": "malint-qwen35-4l-repeated-matched-cv/1.0",
592
- "fold_runner_sha256": "c71af81f1182ce860c39c107dd19eff02f1addd8dd646e3eebaadedc4503067c",
593
- "heldout_split_accessed": false,
594
- "post_selection_internal_evidence": true,
595
- "raw_text_written": false,
596
- "repeat_count": 3,
597
- "rows_per_repeat": 1100,
598
- "terminal_evaluation_only": true,
599
- "threshold": 0.3
600
- }
601
- },
602
- "malint_resources": {
603
- "all_resource_parity_gates_passed": true,
604
- "claim_boundary": {
605
- "development_labels_read": false,
606
- "heldout_access": false,
607
- "quality_scored": false,
608
- "resource_only": true,
609
- "sota_claim_supported": false,
610
- "timing_scope": "CUDA model forward with CUDA-resident token IDs and masks; tokenization, loading, disk, and host-to-device transfer excluded"
611
- },
612
- "comparisons": {
613
- "1": {
614
- "existing_over_mean_kd_throughput_ratio": 1.0456925078172845,
615
- "mean_kd_minus_existing_parameters": 5,
616
- "mean_kd_minus_existing_serialized_weight_bytes": -220,
617
- "mean_kd_over_existing_latency_ratio": 1.0456925078172845,
618
- "mean_kd_over_existing_peak_allocated_ratio": 1.0007640009038254,
619
- "paired_round_latency_ratio_summary": {
620
- "maximum": 1.0527066487704841,
621
- "mean": 1.0488242542281458,
622
- "median": 1.0511611345543765,
623
- "minimum": 1.042604979359577,
624
- "n": 3,
625
- "p05": 1.043460594879057,
626
- "p95": 1.0525520973488733,
627
- "sample_standard_deviation": 0.005441202851039159
628
- },
629
- "paired_round_latency_ratios": [
630
- 1.0511611345543765,
631
- 1.042604979359577,
632
- 1.0527066487704841
633
- ],
634
- "resource_parity_gate": {
635
- "four_layers_and_length_256_pass": true,
636
- "latency_pass": true,
637
- "latency_ratio": 1.0456925078172845,
638
- "latency_ratio_limit": 1.05,
639
- "parameter_pass": true,
640
- "parameter_relative_difference": 1.4965497464629228e-08,
641
- "parameter_relative_limit": 0.0001,
642
- "passed": true,
643
- "peak_allocated_ratio": 1.0007640009038254,
644
- "peak_memory_pass": true,
645
- "peak_memory_ratio_limit": 1.05
646
- }
647
- },
648
- "4": {
649
- "existing_over_mean_kd_throughput_ratio": 1.0177048568315403,
650
- "mean_kd_minus_existing_parameters": 5,
651
- "mean_kd_minus_existing_serialized_weight_bytes": -220,
652
- "mean_kd_over_existing_latency_ratio": 1.0177048568315403,
653
- "mean_kd_over_existing_peak_allocated_ratio": 1.0007213257185181,
654
- "paired_round_latency_ratio_summary": {
655
- "maximum": 1.0378811577131521,
656
- "mean": 1.0299336811276214,
657
- "median": 1.0272510844529572,
658
- "minimum": 1.0246688012167542,
659
- "n": 3,
660
- "p05": 1.0249270295403745,
661
- "p95": 1.0368181503871325,
662
- "sample_standard_deviation": 0.007002773360331318
663
- },
664
- "paired_round_latency_ratios": [
665
- 1.0272510844529572,
666
- 1.0378811577131521,
667
- 1.0246688012167542
668
- ],
669
- "resource_parity_gate": {
670
- "four_layers_and_length_256_pass": true,
671
- "latency_pass": true,
672
- "latency_ratio": 1.0177048568315403,
673
- "latency_ratio_limit": 1.05,
674
- "parameter_pass": true,
675
- "parameter_relative_difference": 1.4965497464629228e-08,
676
- "parameter_relative_limit": 0.0001,
677
- "passed": true,
678
- "peak_allocated_ratio": 1.0007213257185181,
679
- "peak_memory_pass": true,
680
- "peak_memory_ratio_limit": 1.05
681
- }
682
- }
683
- },
684
- "rows": [
685
- {
686
- "arm": "existing_rightmost",
687
- "batch": 1,
688
- "latency_ms_per_example": 18.474629627574572,
689
- "parameter_count": 334101824,
690
- "peak_allocated_bytes": 1373150208,
691
- "worker_count": 3
692
- },
693
- {
694
- "arm": "mean_kd",
695
- "batch": 1,
696
- "latency_ms_per_example": 19.318781786253957,
697
- "parameter_count": 334101829,
698
- "peak_allocated_bytes": 1374199296,
699
- "worker_count": 3
700
- },
701
- {
702
- "arm": "existing_rightmost",
703
- "batch": 4,
704
- "latency_ms_per_example": 4.589751928502863,
705
- "parameter_count": 334101824,
706
- "peak_allocated_bytes": 1454388736,
707
- "worker_count": 3
708
- },
709
- {
710
- "arm": "mean_kd",
711
- "batch": 4,
712
- "latency_ms_per_example": 4.671012829289292,
713
- "parameter_count": 334101829,
714
- "peak_allocated_bytes": 1455437824,
715
- "worker_count": 3
716
- }
717
- ]
718
- },
719
- "privacy": {
720
- "aggregate_only": true,
721
- "contains_article_ids": false,
722
- "contains_logits_probabilities_or_predictions": false,
723
- "contains_raw_text": false,
724
- "contains_row_labels": false
725
- },
726
- "propaganda_operational_endpoints": [
727
- {
728
- "arm": "qwen35-full24l-commonhead",
729
- "label": "24L",
730
- "layers": 24,
731
- "lineage_note": "same five-epoch common-head protocol",
732
- "macro_f1": {
733
- "mean": 0.5642638295960175,
734
- "n": 3,
735
- "sample_sd": 0.006258972649035587
736
- },
737
- "micro_f1": {
738
- "mean": 0.6264844890231888,
739
- "n": 3,
740
- "sample_sd": 0.01398932248989461
741
- },
742
- "parameters": 752407360,
743
- "peak_cuda_allocated_bytes": {
744
- "mean": 7635339264.0,
745
- "n": 3,
746
- "sample_sd": 0.0
747
- },
748
- "subset_exact_match": {
749
- "mean": 0.0303030303030303,
750
- "n": 3,
751
- "sample_sd": 0.027773186030035393
752
- },
753
- "test_articles": 55,
754
- "test_seconds_per_article": {
755
- "mean": 0.586362602903109,
756
- "n": 3,
757
- "sample_sd": 0.005979023788731576
758
- },
759
- "total_wall_seconds": {
760
- "mean": 12842.063612925334,
761
- "n": 3,
762
- "sample_sd": 94.80225434811064
763
- }
764
- },
765
- {
766
- "arm": "qwen35-c8q8l-commonhead",
767
- "label": "8L",
768
- "layers": 8,
769
- "lineage_note": "same five-epoch common-head protocol",
770
- "macro_f1": {
771
- "mean": 0.5366873350531403,
772
- "n": 3,
773
- "sample_sd": 0.017264532146217884
774
- },
775
- "micro_f1": {
776
- "mean": 0.6238830315003392,
777
- "n": 3,
778
- "sample_sd": 0.013623163834245318
779
- },
780
- "parameters": 420332480,
781
- "peak_cuda_allocated_bytes": {
782
- "mean": 4257894400.0,
783
- "n": 3,
784
- "sample_sd": 3632373.815194686
785
- },
786
- "subset_exact_match": {
787
- "mean": 0.006060606060606061,
788
- "n": 3,
789
- "sample_sd": 0.01049727762162956
790
- },
791
- "test_articles": 55,
792
- "test_seconds_per_article": {
793
- "mean": 0.206657676709195,
794
- "n": 3,
795
- "sample_sd": 0.002071265185964238
796
- },
797
- "total_wall_seconds": {
798
- "mean": 4291.582853937677,
799
- "n": 3,
800
- "sample_sd": 44.49609583978631
801
- }
802
- },
803
- {
804
- "arm": "qwen35-standalone4l-commonhead",
805
- "label": "4L",
806
- "layers": 4,
807
- "lineage_note": "same five-epoch common-head protocol",
808
- "macro_f1": {
809
- "mean": 0.5875679559555907,
810
- "n": 3,
811
- "sample_sd": 0.007165753447433348
812
- },
813
- "micro_f1": {
814
- "mean": 0.647978512363484,
815
- "n": 3,
816
- "sample_sd": 0.0018739735696546695
817
- },
818
- "parameters": 210903360,
819
- "peak_cuda_allocated_bytes": {
820
- "mean": 2143655936.0,
821
- "n": 3,
822
- "sample_sd": 0.0
823
- },
824
- "subset_exact_match": {
825
- "mean": 0.006060606060606061,
826
- "n": 3,
827
- "sample_sd": 0.01049727762162956
828
- },
829
- "test_articles": 55,
830
- "test_seconds_per_article": {
831
- "mean": 0.08543555729062946,
832
- "n": 3,
833
- "sample_sd": 0.0005980667826383295
834
- },
835
- "total_wall_seconds": {
836
- "mean": 1643.6388439506602,
837
- "n": 3,
838
- "sample_sd": 9.226529522850198
839
- }
840
- }
841
- ],
842
- "schema": "qwen35-4l-cross-task-base-card-figures-v1",
843
- "status": "complete",
844
- "task_free_compression_ladder": [
845
- {
846
- "layers": 24,
847
- "name": "24L upstream text teacher",
848
- "parameters": 752393024,
849
- "weight_bytes": 1504825632,
850
- "weight_mib": 1435.1135559082031
851
- },
852
- {
853
- "layers": 8,
854
- "name": "8L task-free KD",
855
- "parameters": 420318144,
856
- "weight_bytes": 840647880,
857
- "weight_mib": 801.7042922973633
858
- },
859
- {
860
- "layers": 6,
861
- "name": "6L task-free KD",
862
- "parameters": 377207424,
863
- "weight_bytes": 754423424,
864
- "weight_mib": 719.4742431640625
865
- },
866
- {
867
- "layers": 4,
868
- "name": "4L task-free KD root",
869
- "parameters": 334096704,
870
- "weight_bytes": 668198976,
871
- "weight_mib": 637.2442016601562
872
- }
873
- ],
874
- "uno_q_controlled_depth": {
875
- "boundary": "Matched V128k/F16 runtime comparison. The 4L arm is the historical early-path proxy, not the final common-head quality model.",
876
- "claim_boundaries": [
877
- "Only layer_effect_primary is used for same-vocabulary F16 depth-effect estimates.",
878
- "Only vocabulary_effect_primary is used for exact-remap vocabulary-effect estimates.",
879
- "Operational and cross-generation groups are descriptive because precision, heads, layer maps, training lineages, or architectures differ.",
880
- "llama-bench embedding-mode latency measures backbone/classifier compute but does not validate evidence-span quality.",
881
- "CPU occupancy is instantaneous saturation; core-seconds is a compute-duration proxy, not measured energy.",
882
- "No power meter was used, so energy consumption is not reported.",
883
- "Temperature is device- and ambient-specific and is descriptive, not a universal hardware constant."
884
- ],
885
- "rows": [
886
- {
887
- "layers": 8,
888
- "metric": "artifact_mib",
889
- "metric_label": "Artifact size",
890
- "model": "8L",
891
- "normalized_8l_percent": 100.0
892
- },
893
- {
894
- "layers": 6,
895
- "metric": "artifact_mib",
896
- "metric_label": "Artifact size",
897
- "model": "6L",
898
- "normalized_8l_percent": 85.59176355424236
899
- },
900
- {
901
- "layers": 4,
902
- "metric": "artifact_mib",
903
- "metric_label": "Artifact size",
904
- "model": "4L proxy",
905
- "normalized_8l_percent": 72.2607217304372
906
- },
907
- {
908
- "layers": 8,
909
- "metric": "latency_pp128_ms",
910
- "metric_label": "PP128 latency",
911
- "model": "8L",
912
- "normalized_8l_percent": 100.0
913
- },
914
- {
915
- "layers": 6,
916
- "metric": "latency_pp128_ms",
917
- "metric_label": "PP128 latency",
918
- "model": "6L",
919
- "normalized_8l_percent": 76.36565126547553
920
- },
921
- {
922
- "layers": 4,
923
- "metric": "latency_pp128_ms",
924
- "metric_label": "PP128 latency",
925
- "model": "4L proxy",
926
- "normalized_8l_percent": 50.16216836129361
927
- },
928
- {
929
- "layers": 8,
930
- "metric": "peak_rss_mib",
931
- "metric_label": "Peak RSS",
932
- "model": "8L",
933
- "normalized_8l_percent": 100.0
934
- },
935
- {
936
- "layers": 6,
937
- "metric": "peak_rss_mib",
938
- "metric_label": "Peak RSS",
939
- "model": "6L",
940
- "normalized_8l_percent": 87.50983221571457
941
- },
942
- {
943
- "layers": 4,
944
- "metric": "peak_rss_mib",
945
- "metric_label": "Peak RSS",
946
- "model": "4L proxy",
947
- "normalized_8l_percent": 78.19200407435702
948
- },
949
- {
950
- "layers": 8,
951
- "metric": "mean_core_seconds_per_round",
952
- "metric_label": "Core-seconds",
953
- "model": "8L",
954
- "normalized_8l_percent": 100.0
955
- },
956
- {
957
- "layers": 6,
958
- "metric": "mean_core_seconds_per_round",
959
- "metric_label": "Core-seconds",
960
- "model": "6L",
961
- "normalized_8l_percent": 76.16702179915283
962
- },
963
- {
964
- "layers": 4,
965
- "metric": "mean_core_seconds_per_round",
966
- "metric_label": "Core-seconds",
967
- "model": "4L proxy",
968
- "normalized_8l_percent": 51.03561371028788
969
- }
970
- ]
971
- }
972
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
benchmark/figures/fig10.py ADDED
@@ -0,0 +1,92 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Figure 10: the published root against every arm measured under the frozen protocol, on every metric."""
2
+ import csv, json, pathlib
3
+ import matplotlib
4
+ matplotlib.use("Agg")
5
+ import matplotlib.pyplot as plt
6
+ import numpy as np
7
+
8
+ S = pathlib.Path("/tmp/claude-1000/-home-mp-ubuntu-Projects-paper-qwen-intent-classifier/"
9
+ "0af32b5a-ac18-4e73-a3ad-1c863d0154af/scratchpad")
10
+ OUT = S / "fig2"
11
+ SURFACE = "#fcfcfb"; INK, INK2, MUTED = "#1a1a19", "#4a4a47", "#8a8a85"; GRID = "#e5e5e1"
12
+ C = {"blue": "#2a78d6", "orange": "#eb6834", "aqua": "#1baf7a", "yellow": "#eda100", "violet": "#4a3aa7"}
13
+ plt.rcParams.update({
14
+ "figure.facecolor": SURFACE, "axes.facecolor": SURFACE, "savefig.facecolor": SURFACE,
15
+ "axes.edgecolor": GRID, "axes.linewidth": 0.8, "axes.labelcolor": INK2,
16
+ "xtick.color": MUTED, "ytick.color": MUTED, "text.color": INK,
17
+ "font.size": 9, "axes.titlesize": 10.5, "axes.titleweight": "semibold",
18
+ "grid.color": GRID, "grid.linewidth": 0.7,
19
+ "xtick.major.size": 0, "ytick.major.size": 0,
20
+ "legend.frameon": False, "figure.dpi": 160,
21
+ })
22
+
23
+ arms = json.loads((S / "merged_arms.json").read_text())
24
+ MI = 1024 ** 2
25
+ rows = {r["arm_id"]: r for r in csv.DictReader(open(str(S / "hf/bench/61_arm_bf16.csv")))}
26
+ g = json.loads(pathlib.Path("/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/agents/sessions/"
27
+ "2026-09-20/S-20260920-taskblind-ladder-v1/MEASUREMENTS.json").read_text())
28
+ gridN = {"15,380": "8192", "23,551": "16384", "39,866": "32768", "72,455": "65536"}
29
+ for a in arms:
30
+ if a["new"]:
31
+ n = next(v for k, v in gridN.items() if k in a["arm_id"])
32
+ a["peak_rss_mib"] = g["arms"][n]["BF16"]["peak_tree_rss_bytes"] / MI
33
+ else:
34
+ a["peak_rss_mib"] = float(rows[a["arm_id"]]["peak_rss_mib"])
35
+
36
+ FINAL = "39,866"
37
+ def fam(a):
38
+ if a["new"]: return "final" if FINAL in a["arm_id"] else "grid"
39
+ return "qwen" if a["arm_id"].startswith(("qwen35", "qwen25")) else "encoder"
40
+ COL = {"encoder": MUTED, "qwen": C["violet"], "grid": C["blue"], "final": C["orange"]}
41
+
42
+ PANELS = [
43
+ ("macro_f1_mean", "macro F1 — higher is better", True, False, "{:.4f}"),
44
+ ("p50_ms", "document p50, ms — lower is better", False, True, "{:.1f}"),
45
+ ("peak_rss_mib", "peak process-tree RSS, MiB — lower is better", False, False, "{:.0f}"),
46
+ ("weight_file_mib", "BF16 weight file, MiB — lower is better", False, True, "{:.0f}"),
47
+ ("gpu_j_per_doc", "GPU joules per document — lower is better", False, True, "{:.2f}"),
48
+ ]
49
+
50
+ fig, axes = plt.subplots(len(PANELS), 1, figsize=(11.0, 7.4))
51
+ for ax, (key, title, higher, logx, fmt) in zip(axes, PANELS):
52
+ ax.set_axisbelow(True)
53
+ ax.grid(True, axis="x", alpha=0.9)
54
+ for s in ("top", "right", "left"): ax.spines[s].set_visible(False)
55
+ ax.spines["bottom"].set_color(GRID)
56
+ vals = [a[key] for a in arms]
57
+ root = next(a for a in arms if fam(a) == "final")
58
+ order = sorted(vals, reverse=higher)
59
+ rank = order.index(root[key]) + 1
60
+ for a in arms:
61
+ f = fam(a)
62
+ if f == "final": continue
63
+ ax.plot([a[key], a[key]], [0.18, 0.82], lw=1.6 if f == "grid" else 1.1,
64
+ color=COL[f], alpha=0.95 if f == "grid" else 0.45, solid_capstyle="butt", zorder=3)
65
+ ax.plot([root[key]], [0.5], marker="*", ms=17, color=C["orange"],
66
+ markeredgecolor=SURFACE, markeredgewidth=1.2, zorder=6)
67
+ if logx: ax.set_xscale("log")
68
+ ax.set_yticks([]); ax.set_ylim(0, 1.25)
69
+ ax.set_title(title, loc="left", pad=6)
70
+ good = "top" if rank <= len(arms) / 3 else ("middle" if rank <= 2 * len(arms) / 3 else "bottom")
71
+ ax.annotate(f"this root {fmt.format(root[key])} · rank {rank} of {len(arms)} ({good} third)",
72
+ (0.998, 1.06), xycoords="axes fraction", ha="right", va="bottom",
73
+ fontsize=8.5, color=C["orange"], fontweight="semibold")
74
+ lo, hi = min(vals), max(vals)
75
+ ax.annotate(f"best {fmt.format(order[0])}", (hi if higher else lo, 0.02),
76
+ xycoords=("data", "axes fraction"), ha="right" if higher else "left",
77
+ va="bottom", fontsize=7, color=MUTED)
78
+
79
+ handles = [plt.Line2D([], [], color=COL["encoder"], lw=2, label="encoder families (46)"),
80
+ plt.Line2D([], [], color=COL["qwen"], lw=2, label="earlier Qwen arms (15)"),
81
+ plt.Line2D([], [], color=COL["grid"], lw=2, label="task-blind vocabulary grid (4)"),
82
+ plt.Line2D([], [], color=C["orange"], marker="*", ms=13, lw=0, label="this root: 4L, 39,866 EN/KO")]
83
+ fig.legend(handles=handles, loc="lower center", ncol=4, fontsize=8.5, bbox_to_anchor=(0.5, 0.062))
84
+ fig.suptitle("The published root against all 65 arms, on every metric the protocol measured",
85
+ fontsize=12, fontweight="semibold", y=0.988)
86
+ fig.text(0.5, 0.030, "Each tick is one arm. Quality is a transfer probe with a fresh 14-label head, three seeds; the other four are single measured values.",
87
+ ha="center", fontsize=7.5, color=MUTED)
88
+ fig.text(0.5, 0.008, "The four grid arms were measured in a later session; a same-host check on an earlier cut put the host effect at 1.05x on BF16 latency and 0.99x on RSS, so read small latency gaps with that in mind.",
89
+ ha="center", fontsize=7.5, color=MUTED)
90
+ fig.subplots_adjust(left=0.035, right=0.985, top=0.925, bottom=0.135, hspace=0.72)
91
+ fig.savefig(OUT / "10_root_on_every_metric.png")
92
+ print("wrote 10_root_on_every_metric.png")
benchmark/figures/manifest.json DELETED
@@ -1,14 +0,0 @@
1
- {
2
- "schema_version": "standalone4l-v128k-figures-v1",
3
- "quality_scope": "existing opened SemEval transfer evidence; 128k vs 248k is not causal because lineages differ",
4
- "source_files": [
5
- "results/standalone4l_base/20260731-v1/huggingface_release_v3/benchmark/arm_summary.csv",
6
- "results/standalone4l_base/20260731-v1/huggingface_release_v3/benchmark/compression_ladder.csv",
7
- "results/standalone4l_base/20260731-v1/huggingface_release_v3/benchmark/resource_metrics.csv"
8
- ],
9
- "figures": [
10
- "v128k_vocabulary_size.png",
11
- "v128k_quality_storage.png"
12
- ],
13
- "new_headless_quality": "NR: no new downstream benchmark run"
14
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
benchmark/figures/source_summary.json DELETED
@@ -1,891 +0,0 @@
1
- {
2
- "schema": "standalone4l-complete-transfer-benchmark-v2",
3
- "status": "complete_from_frozen_reports",
4
- "claim_scope": "Exploratory fresh-head transfer evidence on one English task and a previously opened 55-article SemEval-derived test split.",
5
- "headless_root_requires_adaptation": true,
6
- "test_previously_opened": true,
7
- "test_articles": 55,
8
- "test_windows": 434,
9
- "seeds": [
10
- 41,
11
- 42,
12
- 43
13
- ],
14
- "three_seed_sd_is_not_confidence_interval": true,
15
- "multilingual_downstream_evaluation_performed": false,
16
- "upstream_multilingual_metadata": {
17
- "source": "Qwen/Qwen3.5-0.8B official model card",
18
- "languages_and_dialects_claimed_upstream": 201,
19
- "scope": "inherited tokenizer and architecture coverage only"
20
- },
21
- "protocol": {
22
- "epochs": 5,
23
- "primary_epoch": 5,
24
- "early_stopping": false,
25
- "max_length": 256,
26
- "stride": 128,
27
- "window_target": "span-aligned technique labels",
28
- "article_aggregation": "labelwise maximum probability",
29
- "micro_batch_size": 1,
30
- "gradient_accumulation": 32,
31
- "effective_batch_size": 32,
32
- "learning_rate": 2e-05,
33
- "weight_decay": 0.01,
34
- "gradient_clip": 1.0,
35
- "optimizer": "AdamW",
36
- "dtype": "bfloat16",
37
- "thresholds": [
38
- 0.05,
39
- 0.1,
40
- 0.15,
41
- 0.2,
42
- 0.25,
43
- 0.3,
44
- 0.35,
45
- 0.4,
46
- 0.45,
47
- 0.5,
48
- 0.55,
49
- 0.6,
50
- 0.65,
51
- 0.7,
52
- 0.75,
53
- 0.8,
54
- 0.85,
55
- 0.9
56
- ],
57
- "threshold_selection": "epoch-5 calibration macro-F1, then micro-F1"
58
- },
59
- "arms": {
60
- "task_agnostic_base": {
61
- "display": "Task-agnostic KD base",
62
- "lineage": "root base plus a fresh 14-label transfer head",
63
- "metrics": {
64
- "macro_f1": {
65
- "values": [
66
- 0.5883093293517536,
67
- 0.6028379036218096,
68
- 0.5712816197954101
69
- ],
70
- "mean": 0.5874762842563245,
71
- "sample_sd": 0.01579462678670608,
72
- "n": 3
73
- },
74
- "micro_f1": {
75
- "values": [
76
- 0.6468085106382978,
77
- 0.6655172413793103,
78
- 0.6465256797583082
79
- ],
80
- "mean": 0.6529504772586388,
81
- "sample_sd": 0.010884055708931178,
82
- "n": 3
83
- },
84
- "exact_match": {
85
- "values": [
86
- 0.05454545454545454,
87
- 0.03636363636363636,
88
- 0.0
89
- ],
90
- "mean": 0.0303030303030303,
91
- "sample_sd": 0.027773186030035393,
92
- "n": 3
93
- },
94
- "predicted_positive_rate": {
95
- "values": [
96
- 0.5714285714285714,
97
- 0.4090909090909091,
98
- 0.5155844155844156
99
- ],
100
- "mean": 0.4987012987012987,
101
- "sample_sd": 0.08247520160679181,
102
- "n": 3
103
- }
104
- },
105
- "resources": {
106
- "parameters": {
107
- "values": [
108
- 334111040.0,
109
- 334111040.0,
110
- 334111040.0
111
- ],
112
- "mean": 334111040.0,
113
- "sample_sd": 0.0,
114
- "n": 3
115
- },
116
- "peak_cuda_allocated_bytes": {
117
- "values": [
118
- 3380975616.0,
119
- 3380975616.0,
120
- 3380975616.0
121
- ],
122
- "mean": 3380975616.0,
123
- "sample_sd": 0.0,
124
- "n": 3
125
- },
126
- "peak_cuda_reserved_bytes": {
127
- "values": [
128
- 3519021056.0,
129
- 3519021056.0,
130
- 3510632448.0
131
- ],
132
- "mean": 3516224853.3333335,
133
- "sample_sd": 4843165.086926248,
134
- "n": 3
135
- },
136
- "total_wall_seconds": {
137
- "values": [
138
- 1583.0297977919981,
139
- 1588.617994428001,
140
- 1603.4310500120046
141
- ],
142
- "mean": 1591.692947410668,
143
- "sample_sd": 10.54249851412311,
144
- "n": 3
145
- },
146
- "test_seconds_per_article": {
147
- "values": [
148
- 0.08164824154545468,
149
- 0.07871871667289683,
150
- 0.08355205721825107
151
- ],
152
- "mean": 0.08130633847886752,
153
- "sample_sd": 0.0024347419754560395,
154
- "n": 3
155
- }
156
- },
157
- "thresholds": [
158
- 0.15,
159
- 0.3,
160
- 0.1
161
- ]
162
- },
163
- "structural_copy_control": {
164
- "display": "Structural copy, no task-free KD",
165
- "lineage": "same four-layer structure without task-free KD",
166
- "metrics": {
167
- "macro_f1": {
168
- "values": [
169
- 0.5169424579262757,
170
- 0.5349405779540175,
171
- 0.5023778364496152
172
- ],
173
- "mean": 0.5180869574433028,
174
- "sample_sd": 0.016311512588628545,
175
- "n": 3
176
- },
177
- "micro_f1": {
178
- "values": [
179
- 0.5773955773955775,
180
- 0.6083445491251682,
181
- 0.585635359116022
182
- ],
183
- "mean": 0.5904584952122559,
184
- "sample_sd": 0.01602830914762526,
185
- "n": 3
186
- },
187
- "exact_match": {
188
- "values": [
189
- 0.0,
190
- 0.0,
191
- 0.0
192
- ],
193
- "mean": 0.0,
194
- "sample_sd": 0.0,
195
- "n": 3
196
- },
197
- "predicted_positive_rate": {
198
- "values": [
199
- 0.712987012987013,
200
- 0.6207792207792208,
201
- 0.5961038961038961
202
- ],
203
- "mean": 0.6432900432900432,
204
- "sample_sd": 0.06160737456395461,
205
- "n": 3
206
- }
207
- },
208
- "resources": {
209
- "parameters": {
210
- "values": [
211
- 334111040.0,
212
- 334111040.0,
213
- 334111040.0
214
- ],
215
- "mean": 334111040.0,
216
- "sample_sd": 0.0,
217
- "n": 3
218
- },
219
- "peak_cuda_allocated_bytes": {
220
- "values": [
221
- 3380975616.0,
222
- 3380975616.0,
223
- 3380975616.0
224
- ],
225
- "mean": 3380975616.0,
226
- "sample_sd": 0.0,
227
- "n": 3
228
- },
229
- "peak_cuda_reserved_bytes": {
230
- "values": [
231
- 3519021056.0,
232
- 3519021056.0,
233
- 3510632448.0
234
- ],
235
- "mean": 3516224853.3333335,
236
- "sample_sd": 4843165.086926248,
237
- "n": 3
238
- },
239
- "total_wall_seconds": {
240
- "values": [
241
- 1575.5905925639963,
242
- 1623.5639530190092,
243
- 1629.195114912989
244
- ],
245
- "mean": 1609.4498868319981,
246
- "sample_sd": 29.457874198835384,
247
- "n": 3
248
- },
249
- "test_seconds_per_article": {
250
- "values": [
251
- 0.07877608058201573,
252
- 0.08307423260041767,
253
- 0.08163602005436339
254
- ],
255
- "mean": 0.08116211107893226,
256
- "sample_sd": 0.002187914527825257,
257
- "n": 3
258
- }
259
- },
260
- "thresholds": [
261
- 0.05,
262
- 0.2,
263
- 0.2
264
- ]
265
- },
266
- "existing_specialized_separate_lineage": {
267
- "display": "Existing specialized 4L, separate lineage",
268
- "lineage": "historical task-specific shrink/distillation lineage",
269
- "metrics": {
270
- "macro_f1": {
271
- "values": [
272
- 0.5832848452288824,
273
- 0.5958405181160762,
274
- 0.5835785045218137
275
- ],
276
- "mean": 0.5875679559555907,
277
- "sample_sd": 0.007165753447433348,
278
- "n": 3
279
- },
280
- "micro_f1": {
281
- "values": [
282
- 0.6494252873563218,
283
- 0.6458616010854816,
284
- 0.6486486486486486
285
- ],
286
- "mean": 0.647978512363484,
287
- "sample_sd": 0.0018739735696546695,
288
- "n": 3
289
- },
290
- "exact_match": {
291
- "values": [
292
- 0.0,
293
- 0.0,
294
- 0.01818181818181818
295
- ],
296
- "mean": 0.006060606060606061,
297
- "sample_sd": 0.01049727762162956,
298
- "n": 3
299
- },
300
- "predicted_positive_rate": {
301
- "values": [
302
- 0.5597402597402598,
303
- 0.612987012987013,
304
- 0.5688311688311688
305
- ],
306
- "mean": 0.5805194805194805,
307
- "sample_sd": 0.02848274311618351,
308
- "n": 3
309
- }
310
- },
311
- "resources": {
312
- "parameters": {
313
- "values": [
314
- 210903360.0,
315
- 210903360.0,
316
- 210903360.0
317
- ],
318
- "mean": 210903360.0,
319
- "sample_sd": 0.0,
320
- "n": 3
321
- },
322
- "peak_cuda_allocated_bytes": {
323
- "values": [
324
- 2143655936.0,
325
- 2143655936.0,
326
- 2143655936.0
327
- ],
328
- "mean": 2143655936.0,
329
- "sample_sd": 0.0,
330
- "n": 3
331
- },
332
- "peak_cuda_reserved_bytes": {
333
- "values": [
334
- 2281701376.0,
335
- 2281701376.0,
336
- 2273312768.0
337
- ],
338
- "mean": 2278905173.3333335,
339
- "sample_sd": 4843165.086926248,
340
- "n": 3
341
- },
342
- "total_wall_seconds": {
343
- "values": [
344
- 1634.488557530014,
345
- 1652.9397718419787,
346
- 1643.488202479988
347
- ],
348
- "mean": 1643.6388439506602,
349
- "sample_sd": 9.226529522850198,
350
- "n": 3
351
- },
352
- "test_seconds_per_article": {
353
- "values": [
354
- 0.0847959352360869,
355
- 0.08598086021798239,
356
- 0.08552987641781908
357
- ],
358
- "mean": 0.08543555729062946,
359
- "sample_sd": 0.0005980667826383295,
360
- "n": 3
361
- }
362
- },
363
- "thresholds": [
364
- 0.1,
365
- 0.05,
366
- 0.05
367
- ]
368
- }
369
- },
370
- "same_seed_macro_f1_deltas_base_minus_control": [
371
- 0.07136687142547793,
372
- 0.06789732566779205,
373
- 0.06890378334579494
374
- ],
375
- "mean_macro_f1_delta_base_minus_control": 0.06938932681302164,
376
- "publication_gate_passed": true,
377
- "per_label_summary": [
378
- {
379
- "arm": "task_agnostic_base",
380
- "label": "Appeal_to_Authority",
381
- "support": 11,
382
- "precision_mean": 0.3958614484930274,
383
- "precision_sample_sd": 0.1238064457010119,
384
- "recall_mean": 0.5757575757575757,
385
- "recall_sample_sd": 0.05248638810814781,
386
- "f1_mean": 0.4657088122605364,
387
- "f1_sample_sd": 0.1020990126979069
388
- },
389
- {
390
- "arm": "task_agnostic_base",
391
- "label": "Appeal_to_fear-prejudice",
392
- "support": 24,
393
- "precision_mean": 0.5540244550770866,
394
- "precision_sample_sd": 0.02525673990402998,
395
- "recall_mean": 0.7361111111111112,
396
- "recall_sample_sd": 0.13393959390267993,
397
- "f1_mean": 0.6277521761392729,
398
- "f1_sample_sd": 0.04994879016183189
399
- },
400
- {
401
- "arm": "task_agnostic_base",
402
- "label": "Bandwagon,Reductio_ad_hitlerum",
403
- "support": 7,
404
- "precision_mean": 0.4851851851851852,
405
- "precision_sample_sd": 0.07882887649552875,
406
- "recall_mean": 0.8095238095238095,
407
- "recall_sample_sd": 0.08247860988423221,
408
- "f1_mean": 0.6006778309409889,
409
- "f1_sample_sd": 0.04793776258918466
410
- },
411
- {
412
- "arm": "task_agnostic_base",
413
- "label": "Black-and-White_Fallacy",
414
- "support": 12,
415
- "precision_mean": 0.3667929292929293,
416
- "precision_sample_sd": 0.07670649392711647,
417
- "recall_mean": 0.47222222222222227,
418
- "recall_sample_sd": 0.09622504486493764,
419
- "f1_mean": 0.4053892966936445,
420
- "f1_sample_sd": 0.04211369218048746
421
- },
422
- {
423
- "arm": "task_agnostic_base",
424
- "label": "Causal_Oversimplification",
425
- "support": 18,
426
- "precision_mean": 0.5428824049513704,
427
- "precision_sample_sd": 0.039216350192122156,
428
- "recall_mean": 0.7962962962962963,
429
- "recall_sample_sd": 0.11564811108145183,
430
- "f1_mean": 0.6446992723588468,
431
- "f1_sample_sd": 0.0634561308339175
432
- },
433
- {
434
- "arm": "task_agnostic_base",
435
- "label": "Doubt",
436
- "support": 22,
437
- "precision_mean": 0.5223832528180354,
438
- "precision_sample_sd": 0.06742780050556177,
439
- "recall_mean": 0.9848484848484849,
440
- "recall_sample_sd": 0.026243194054073875,
441
- "f1_mean": 0.6802057822343391,
442
- "f1_sample_sd": 0.04928561754686331
443
- },
444
- {
445
- "arm": "task_agnostic_base",
446
- "label": "Exaggeration,Minimisation",
447
- "support": 22,
448
- "precision_mean": 0.4620811287477954,
449
- "precision_sample_sd": 0.016801198309751,
450
- "recall_mean": 0.7727272727272727,
451
- "recall_sample_sd": 0.15745916432444337,
452
- "f1_mean": 0.5727040816326531,
453
- "f1_sample_sd": 0.03645259990419198
454
- },
455
- {
456
- "arm": "task_agnostic_base",
457
- "label": "Flag-Waving",
458
- "support": 17,
459
- "precision_mean": 0.5642642642642642,
460
- "precision_sample_sd": 0.1478145677656334,
461
- "recall_mean": 0.8235294117647058,
462
- "recall_sample_sd": 0.1764705882352941,
463
- "f1_mean": 0.6464506172839506,
464
- "f1_sample_sd": 0.035742219952758225
465
- },
466
- {
467
- "arm": "task_agnostic_base",
468
- "label": "Loaded_Language",
469
- "support": 45,
470
- "precision_mean": 0.8649032907793016,
471
- "precision_sample_sd": 0.046984312428427724,
472
- "recall_mean": 0.9851851851851852,
473
- "recall_sample_sd": 0.012830005981991702,
474
- "f1_mean": 0.9208741623431562,
475
- "f1_sample_sd": 0.03200926878737141
476
- },
477
- {
478
- "arm": "task_agnostic_base",
479
- "label": "Name_Calling,Labeling",
480
- "support": 33,
481
- "precision_mean": 0.7380952380952381,
482
- "precision_sample_sd": 0.0047619047619048005,
483
- "recall_mean": 0.9090909090909091,
484
- "recall_sample_sd": 0.10925912955951485,
485
- "f1_mean": 0.8125087983911513,
486
- "f1_sample_sd": 0.04252971464057227
487
- },
488
- {
489
- "arm": "task_agnostic_base",
490
- "label": "Repetition",
491
- "support": 23,
492
- "precision_mean": 0.5196620583717357,
493
- "precision_sample_sd": 0.021645909050194662,
494
- "recall_mean": 0.7536231884057971,
495
- "recall_sample_sd": 0.06641414050660639,
496
- "f1_mean": 0.6147520781519571,
497
- "f1_sample_sd": 0.03505927109416221
498
- },
499
- {
500
- "arm": "task_agnostic_base",
501
- "label": "Slogans",
502
- "support": 14,
503
- "precision_mean": 0.5161361626878869,
504
- "precision_sample_sd": 0.20305677930281646,
505
- "recall_mean": 0.7380952380952381,
506
- "recall_sample_sd": 0.10910894511799614,
507
- "f1_mean": 0.5834824090638044,
508
- "f1_sample_sd": 0.09862686289220693
509
- },
510
- {
511
- "arm": "task_agnostic_base",
512
- "label": "Thought-terminating_Cliches",
513
- "support": 7,
514
- "precision_mean": 0.3492063492063492,
515
- "precision_sample_sd": 0.143516994603374,
516
- "recall_mean": 0.3333333333333333,
517
- "recall_sample_sd": 0.08247860988423225,
518
- "f1_mean": 0.319014319014319,
519
- "f1_sample_sd": 0.04017590711117064
520
- },
521
- {
522
- "arm": "task_agnostic_base",
523
- "label": "Whataboutism,Straw_Men,Red_Herring",
524
- "support": 10,
525
- "precision_mean": 0.2761904761904762,
526
- "precision_sample_sd": 0.05302632726504782,
527
- "recall_mean": 0.5,
528
- "recall_sample_sd": 0.2645751311064591,
529
- "f1_mean": 0.330448343079922,
530
- "f1_sample_sd": 0.021845164720973014
531
- },
532
- {
533
- "arm": "structural_copy_control",
534
- "label": "Appeal_to_Authority",
535
- "support": 11,
536
- "precision_mean": 0.26003086419753085,
537
- "precision_sample_sd": 0.06281357095350094,
538
- "recall_mean": 0.5757575757575758,
539
- "recall_sample_sd": 0.262431940540739,
540
- "f1_mean": 0.35477582846003897,
541
- "f1_sample_sd": 0.11479479036518875
542
- },
543
- {
544
- "arm": "structural_copy_control",
545
- "label": "Appeal_to_fear-prejudice",
546
- "support": 24,
547
- "precision_mean": 0.5104813315339631,
548
- "precision_sample_sd": 0.038427508996734555,
549
- "recall_mean": 0.8611111111111112,
550
- "recall_sample_sd": 0.024056261216234387,
551
- "f1_mean": 0.6401745363877202,
552
- "f1_sample_sd": 0.029305467347150463
553
- },
554
- {
555
- "arm": "structural_copy_control",
556
- "label": "Bandwagon,Reductio_ad_hitlerum",
557
- "support": 7,
558
- "precision_mean": 0.2876623376623376,
559
- "precision_sample_sd": 0.061386823077618426,
560
- "recall_mean": 0.8571428571428571,
561
- "recall_sample_sd": 0.14285714285714285,
562
- "f1_mean": 0.4306391777656145,
563
- "f1_sample_sd": 0.08686392624022733
564
- },
565
- {
566
- "arm": "structural_copy_control",
567
- "label": "Black-and-White_Fallacy",
568
- "support": 12,
569
- "precision_mean": 0.31339031339031337,
570
- "precision_sample_sd": 0.08301311842925053,
571
- "recall_mean": 0.5277777777777778,
572
- "recall_sample_sd": 0.2678791878053599,
573
- "f1_mean": 0.38778427550357375,
574
- "f1_sample_sd": 0.13069755090515112
575
- },
576
- {
577
- "arm": "structural_copy_control",
578
- "label": "Causal_Oversimplification",
579
- "support": 18,
580
- "precision_mean": 0.4464201043148412,
581
- "precision_sample_sd": 0.013538446216471536,
582
- "recall_mean": 0.9259259259259259,
583
- "recall_sample_sd": 0.03207501495497923,
584
- "f1_mean": 0.6023809523809524,
585
- "f1_sample_sd": 0.018643640071557464
586
- },
587
- {
588
- "arm": "structural_copy_control",
589
- "label": "Doubt",
590
- "support": 22,
591
- "precision_mean": 0.4455314009661836,
592
- "precision_sample_sd": 0.04067072694857813,
593
- "recall_mean": 1.0,
594
- "recall_sample_sd": 0.0,
595
- "f1_mean": 0.615686274509804,
596
- "f1_sample_sd": 0.039427200694482675
597
- },
598
- {
599
- "arm": "structural_copy_control",
600
- "label": "Exaggeration,Minimisation",
601
- "support": 22,
602
- "precision_mean": 0.4364176150420584,
603
- "precision_sample_sd": 0.025093958106097894,
604
- "recall_mean": 0.8787878787878788,
605
- "recall_sample_sd": 0.05248638810814775,
606
- "f1_mean": 0.5829803047194351,
607
- "f1_sample_sd": 0.030899288356345502
608
- },
609
- {
610
- "arm": "structural_copy_control",
611
- "label": "Flag-Waving",
612
- "support": 17,
613
- "precision_mean": 0.3814814814814815,
614
- "precision_sample_sd": 0.05010277503136551,
615
- "recall_mean": 0.803921568627451,
616
- "recall_sample_sd": 0.17970885078258195,
617
- "f1_mean": 0.5138595287119652,
618
- "f1_sample_sd": 0.06400931989326974
619
- },
620
- {
621
- "arm": "structural_copy_control",
622
- "label": "Loaded_Language",
623
- "support": 45,
624
- "precision_mean": 0.8492581841638446,
625
- "precision_sample_sd": 0.016026591848331686,
626
- "recall_mean": 1.0,
627
- "recall_sample_sd": 0.0,
628
- "f1_mean": 0.9184311025253588,
629
- "f1_sample_sd": 0.009372233868132662
630
- },
631
- {
632
- "arm": "structural_copy_control",
633
- "label": "Name_Calling,Labeling",
634
- "support": 33,
635
- "precision_mean": 0.6375,
636
- "precision_sample_sd": 0.03410563654946854,
637
- "recall_mean": 0.9696969696969697,
638
- "recall_sample_sd": 0.030303030303030276,
639
- "f1_mean": 0.7685185185185185,
640
- "f1_sample_sd": 0.02023900779105559
641
- },
642
- {
643
- "arm": "structural_copy_control",
644
- "label": "Repetition",
645
- "support": 23,
646
- "precision_mean": 0.44863933452168747,
647
- "precision_sample_sd": 0.01520014564765896,
648
- "recall_mean": 0.9420289855072463,
649
- "recall_sample_sd": 0.06641414050660639,
650
- "f1_mean": 0.607248835423033,
651
- "f1_sample_sd": 0.019858627551810427
652
- },
653
- {
654
- "arm": "structural_copy_control",
655
- "label": "Slogans",
656
- "support": 14,
657
- "precision_mean": 0.3274318274318274,
658
- "precision_sample_sd": 0.05434598572773684,
659
- "recall_mean": 0.4523809523809524,
660
- "recall_sample_sd": 0.1486904285332952,
661
- "f1_mean": 0.3715828924162258,
662
- "f1_sample_sd": 0.07106660236343577
663
- },
664
- {
665
- "arm": "structural_copy_control",
666
- "label": "Thought-terminating_Cliches",
667
- "support": 7,
668
- "precision_mean": 0.1142156862745098,
669
- "precision_sample_sd": 0.012848374923608973,
670
- "recall_mean": 0.38095238095238093,
671
- "recall_sample_sd": 0.1649572197684645,
672
- "f1_mean": 0.17331433998100665,
673
- "f1_sample_sd": 0.029065880233243847
674
- },
675
- {
676
- "arm": "structural_copy_control",
677
- "label": "Whataboutism,Straw_Men,Red_Herring",
678
- "support": 10,
679
- "precision_mean": 0.19344598054275472,
680
- "precision_sample_sd": 0.03373872471827226,
681
- "recall_mean": 0.5666666666666667,
682
- "recall_sample_sd": 0.2081665999466133,
683
- "f1_mean": 0.28584083690299267,
684
- "f1_sample_sd": 0.06078854830133676
685
- },
686
- {
687
- "arm": "existing_specialized_separate_lineage",
688
- "label": "Appeal_to_Authority",
689
- "support": 11,
690
- "precision_mean": 0.37836021505376344,
691
- "precision_sample_sd": 0.10546365636105155,
692
- "recall_mean": 0.6666666666666666,
693
- "recall_sample_sd": 0.2287828616748712,
694
- "f1_mean": 0.46885361552028215,
695
- "f1_sample_sd": 0.09502747629890008
696
- },
697
- {
698
- "arm": "existing_specialized_separate_lineage",
699
- "label": "Appeal_to_fear-prejudice",
700
- "support": 24,
701
- "precision_mean": 0.525708635464733,
702
- "precision_sample_sd": 0.02235198659615819,
703
- "recall_mean": 0.8888888888888888,
704
- "recall_sample_sd": 0.048112522432468774,
705
- "f1_mean": 0.6599065351235072,
706
- "f1_sample_sd": 0.015362381542320084
707
- },
708
- {
709
- "arm": "existing_specialized_separate_lineage",
710
- "label": "Bandwagon,Reductio_ad_hitlerum",
711
- "support": 7,
712
- "precision_mean": 0.44485144485144484,
713
- "precision_sample_sd": 0.016487284575115283,
714
- "recall_mean": 0.7619047619047619,
715
- "recall_sample_sd": 0.1649572197684645,
716
- "f1_mean": 0.5571428571428572,
717
- "f1_sample_sd": 0.051507875363771265
718
- },
719
- {
720
- "arm": "existing_specialized_separate_lineage",
721
- "label": "Black-and-White_Fallacy",
722
- "support": 12,
723
- "precision_mean": 0.4662644544997486,
724
- "precision_sample_sd": 0.004538274146363688,
725
- "recall_mean": 0.5833333333333334,
726
- "recall_sample_sd": 0.08333333333333331,
727
- "f1_mean": 0.516747552149851,
728
- "f1_sample_sd": 0.03589484965697345
729
- },
730
- {
731
- "arm": "existing_specialized_separate_lineage",
732
- "label": "Causal_Oversimplification",
733
- "support": 18,
734
- "precision_mean": 0.4472934472934473,
735
- "precision_sample_sd": 0.013055771210700413,
736
- "recall_mean": 0.8333333333333333,
737
- "recall_sample_sd": 0.1469861839480328,
738
- "f1_mean": 0.5781794553724379,
739
- "f1_sample_sd": 0.028407563015813122
740
- },
741
- {
742
- "arm": "existing_specialized_separate_lineage",
743
- "label": "Doubt",
744
- "support": 22,
745
- "precision_mean": 0.5045661001171619,
746
- "precision_sample_sd": 0.023581174969843577,
747
- "recall_mean": 1.0,
748
- "recall_sample_sd": 0.0,
749
- "f1_mean": 0.6704939668174963,
750
- "f1_sample_sd": 0.020973132363659372
751
- },
752
- {
753
- "arm": "existing_specialized_separate_lineage",
754
- "label": "Exaggeration,Minimisation",
755
- "support": 22,
756
- "precision_mean": 0.44589793281653745,
757
- "precision_sample_sd": 0.010987847993006481,
758
- "recall_mean": 0.9393939393939394,
759
- "recall_sample_sd": 0.06943296507508849,
760
- "f1_mean": 0.6043956043956044,
761
- "f1_sample_sd": 0.022305256187015846
762
- },
763
- {
764
- "arm": "existing_specialized_separate_lineage",
765
- "label": "Flag-Waving",
766
- "support": 17,
767
- "precision_mean": 0.5576388888888889,
768
- "precision_sample_sd": 0.03705367103320583,
769
- "recall_mean": 0.8823529411764706,
770
- "recall_sample_sd": 0.11764705882352944,
771
- "f1_mean": 0.680769868923179,
772
- "f1_sample_sd": 0.04164658769406446
773
- },
774
- {
775
- "arm": "existing_specialized_separate_lineage",
776
- "label": "Loaded_Language",
777
- "support": 45,
778
- "precision_mean": 0.8387650085763294,
779
- "precision_sample_sd": 0.01782556577257849,
780
- "recall_mean": 1.0,
781
- "recall_sample_sd": 0.0,
782
- "f1_mean": 0.9122448979591837,
783
- "f1_sample_sd": 0.01060439269940129
784
- },
785
- {
786
- "arm": "existing_specialized_separate_lineage",
787
- "label": "Name_Calling,Labeling",
788
- "support": 33,
789
- "precision_mean": 0.6957104190221673,
790
- "precision_sample_sd": 0.019201490809269257,
791
- "recall_mean": 0.9696969696969697,
792
- "recall_sample_sd": 0.030303030303030276,
793
- "f1_mean": 0.8101049442821594,
794
- "f1_sample_sd": 0.022092733328802537
795
- },
796
- {
797
- "arm": "existing_specialized_separate_lineage",
798
- "label": "Repetition",
799
- "support": 23,
800
- "precision_mean": 0.4781557067271353,
801
- "precision_sample_sd": 0.009897987322673213,
802
- "recall_mean": 0.9420289855072463,
803
- "recall_sample_sd": 0.06641414050660639,
804
- "f1_mean": 0.6337774426009719,
805
- "f1_sample_sd": 0.016444118333485254
806
- },
807
- {
808
- "arm": "existing_specialized_separate_lineage",
809
- "label": "Slogans",
810
- "support": 14,
811
- "precision_mean": 0.4334215167548501,
812
- "precision_sample_sd": 0.05811581861282867,
813
- "recall_mean": 0.6904761904761905,
814
- "recall_sample_sd": 0.1649572197684645,
815
- "f1_mean": 0.5201316298877274,
816
- "f1_sample_sd": 0.018567913895656114
817
- },
818
- {
819
- "arm": "existing_specialized_separate_lineage",
820
- "label": "Thought-terminating_Cliches",
821
- "support": 7,
822
- "precision_mean": 0.18446115288220552,
823
- "precision_sample_sd": 0.03641251252864963,
824
- "recall_mean": 0.6666666666666666,
825
- "recall_sample_sd": 0.1649572197684645,
826
- "f1_mean": 0.2868626868626869,
827
- "f1_sample_sd": 0.0511620864429468
828
- },
829
- {
830
- "arm": "existing_specialized_separate_lineage",
831
- "label": "Whataboutism,Straw_Men,Red_Herring",
832
- "support": 10,
833
- "precision_mean": 0.2777777777777778,
834
- "precision_sample_sd": 0.0481125224324688,
835
- "recall_mean": 0.4,
836
- "recall_sample_sd": 0.0,
837
- "f1_mean": 0.32634032634032634,
838
- "f1_sample_sd": 0.03229931575886014
839
- }
840
- ],
841
- "compression_ladder": [
842
- {
843
- "name": "24L upstream text teacher",
844
- "layers": 24,
845
- "parameters": 752393024,
846
- "weight_bytes": 1504825632,
847
- "weight_mib": 1435.1135559082031,
848
- "weight_sha256": "30783da4651259c26b8a204821ea426b9526431e0c14da0d1c6fa284e7414e5e",
849
- "kd_stage": "",
850
- "mean_total_loss": "",
851
- "mean_interface_loss": "",
852
- "mean_final_loss": ""
853
- },
854
- {
855
- "name": "8L task-free KD",
856
- "layers": 8,
857
- "parameters": 420318144,
858
- "weight_bytes": 840647880,
859
- "weight_mib": 801.7042922973633,
860
- "weight_sha256": "086fbaf9a4838ecde56cd3541c9c3212b22d0c5ee745cf2c1c5e0bb31e0ec22c",
861
- "kd_stage": "24to8",
862
- "mean_total_loss": 0.17622653172702485,
863
- "mean_interface_loss": 0.21562811763578793,
864
- "mean_final_loss": 0.16626427527853593
865
- },
866
- {
867
- "name": "6L task-free KD",
868
- "layers": 6,
869
- "parameters": 377207424,
870
- "weight_bytes": 754423424,
871
- "weight_mib": 719.4742431640625,
872
- "weight_sha256": "492749c20f77a42f6bc993f0927a82217b94f872b7eacf25612f7520c9bc28ce",
873
- "kd_stage": "8to6",
874
- "mean_total_loss": 0.03382104352249371,
875
- "mean_interface_loss": 0.04486069045515251,
876
- "mean_final_loss": 0.031147060785087888
877
- },
878
- {
879
- "name": "4L task-free KD root",
880
- "layers": 4,
881
- "parameters": 334096704,
882
- "weight_bytes": 668198976,
883
- "weight_mib": 637.2442016601562,
884
- "weight_sha256": "2732c616772fe320cdea228ab4554981418b1b2bf615c4183fb1ac8e6e2168d3",
885
- "kd_stage": "6to4",
886
- "mean_total_loss": 0.047020394468859195,
887
- "mean_interface_loss": 0.07048421185527332,
888
- "mean_final_loss": 0.043119939065377366
889
- }
890
- ]
891
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
benchmark/semeval-transfer-summary.json DELETED
@@ -1,891 +0,0 @@
1
- {
2
- "schema": "standalone4l-complete-transfer-benchmark-v2",
3
- "status": "complete_from_frozen_reports",
4
- "claim_scope": "Exploratory fresh-head transfer evidence on one English task and a previously opened 55-article SemEval-derived test split.",
5
- "headless_root_requires_adaptation": true,
6
- "test_previously_opened": true,
7
- "test_articles": 55,
8
- "test_windows": 434,
9
- "seeds": [
10
- 41,
11
- 42,
12
- 43
13
- ],
14
- "three_seed_sd_is_not_confidence_interval": true,
15
- "multilingual_downstream_evaluation_performed": false,
16
- "upstream_multilingual_metadata": {
17
- "source": "Qwen/Qwen3.5-0.8B official model card",
18
- "languages_and_dialects_claimed_upstream": 201,
19
- "scope": "inherited tokenizer and architecture coverage only"
20
- },
21
- "protocol": {
22
- "epochs": 5,
23
- "primary_epoch": 5,
24
- "early_stopping": false,
25
- "max_length": 256,
26
- "stride": 128,
27
- "window_target": "span-aligned technique labels",
28
- "article_aggregation": "labelwise maximum probability",
29
- "micro_batch_size": 1,
30
- "gradient_accumulation": 32,
31
- "effective_batch_size": 32,
32
- "learning_rate": 2e-05,
33
- "weight_decay": 0.01,
34
- "gradient_clip": 1.0,
35
- "optimizer": "AdamW",
36
- "dtype": "bfloat16",
37
- "thresholds": [
38
- 0.05,
39
- 0.1,
40
- 0.15,
41
- 0.2,
42
- 0.25,
43
- 0.3,
44
- 0.35,
45
- 0.4,
46
- 0.45,
47
- 0.5,
48
- 0.55,
49
- 0.6,
50
- 0.65,
51
- 0.7,
52
- 0.75,
53
- 0.8,
54
- 0.85,
55
- 0.9
56
- ],
57
- "threshold_selection": "epoch-5 calibration macro-F1, then micro-F1"
58
- },
59
- "arms": {
60
- "task_agnostic_base": {
61
- "display": "Task-agnostic KD base",
62
- "lineage": "root base plus a fresh 14-label transfer head",
63
- "metrics": {
64
- "macro_f1": {
65
- "values": [
66
- 0.5883093293517536,
67
- 0.6028379036218096,
68
- 0.5712816197954101
69
- ],
70
- "mean": 0.5874762842563245,
71
- "sample_sd": 0.01579462678670608,
72
- "n": 3
73
- },
74
- "micro_f1": {
75
- "values": [
76
- 0.6468085106382978,
77
- 0.6655172413793103,
78
- 0.6465256797583082
79
- ],
80
- "mean": 0.6529504772586388,
81
- "sample_sd": 0.010884055708931178,
82
- "n": 3
83
- },
84
- "exact_match": {
85
- "values": [
86
- 0.05454545454545454,
87
- 0.03636363636363636,
88
- 0.0
89
- ],
90
- "mean": 0.0303030303030303,
91
- "sample_sd": 0.027773186030035393,
92
- "n": 3
93
- },
94
- "predicted_positive_rate": {
95
- "values": [
96
- 0.5714285714285714,
97
- 0.4090909090909091,
98
- 0.5155844155844156
99
- ],
100
- "mean": 0.4987012987012987,
101
- "sample_sd": 0.08247520160679181,
102
- "n": 3
103
- }
104
- },
105
- "resources": {
106
- "parameters": {
107
- "values": [
108
- 334111040.0,
109
- 334111040.0,
110
- 334111040.0
111
- ],
112
- "mean": 334111040.0,
113
- "sample_sd": 0.0,
114
- "n": 3
115
- },
116
- "peak_cuda_allocated_bytes": {
117
- "values": [
118
- 3380975616.0,
119
- 3380975616.0,
120
- 3380975616.0
121
- ],
122
- "mean": 3380975616.0,
123
- "sample_sd": 0.0,
124
- "n": 3
125
- },
126
- "peak_cuda_reserved_bytes": {
127
- "values": [
128
- 3519021056.0,
129
- 3519021056.0,
130
- 3510632448.0
131
- ],
132
- "mean": 3516224853.3333335,
133
- "sample_sd": 4843165.086926248,
134
- "n": 3
135
- },
136
- "total_wall_seconds": {
137
- "values": [
138
- 1583.0297977919981,
139
- 1588.617994428001,
140
- 1603.4310500120046
141
- ],
142
- "mean": 1591.692947410668,
143
- "sample_sd": 10.54249851412311,
144
- "n": 3
145
- },
146
- "test_seconds_per_article": {
147
- "values": [
148
- 0.08164824154545468,
149
- 0.07871871667289683,
150
- 0.08355205721825107
151
- ],
152
- "mean": 0.08130633847886752,
153
- "sample_sd": 0.0024347419754560395,
154
- "n": 3
155
- }
156
- },
157
- "thresholds": [
158
- 0.15,
159
- 0.3,
160
- 0.1
161
- ]
162
- },
163
- "structural_copy_control": {
164
- "display": "Structural copy, no task-free KD",
165
- "lineage": "same four-layer structure without task-free KD",
166
- "metrics": {
167
- "macro_f1": {
168
- "values": [
169
- 0.5169424579262757,
170
- 0.5349405779540175,
171
- 0.5023778364496152
172
- ],
173
- "mean": 0.5180869574433028,
174
- "sample_sd": 0.016311512588628545,
175
- "n": 3
176
- },
177
- "micro_f1": {
178
- "values": [
179
- 0.5773955773955775,
180
- 0.6083445491251682,
181
- 0.585635359116022
182
- ],
183
- "mean": 0.5904584952122559,
184
- "sample_sd": 0.01602830914762526,
185
- "n": 3
186
- },
187
- "exact_match": {
188
- "values": [
189
- 0.0,
190
- 0.0,
191
- 0.0
192
- ],
193
- "mean": 0.0,
194
- "sample_sd": 0.0,
195
- "n": 3
196
- },
197
- "predicted_positive_rate": {
198
- "values": [
199
- 0.712987012987013,
200
- 0.6207792207792208,
201
- 0.5961038961038961
202
- ],
203
- "mean": 0.6432900432900432,
204
- "sample_sd": 0.06160737456395461,
205
- "n": 3
206
- }
207
- },
208
- "resources": {
209
- "parameters": {
210
- "values": [
211
- 334111040.0,
212
- 334111040.0,
213
- 334111040.0
214
- ],
215
- "mean": 334111040.0,
216
- "sample_sd": 0.0,
217
- "n": 3
218
- },
219
- "peak_cuda_allocated_bytes": {
220
- "values": [
221
- 3380975616.0,
222
- 3380975616.0,
223
- 3380975616.0
224
- ],
225
- "mean": 3380975616.0,
226
- "sample_sd": 0.0,
227
- "n": 3
228
- },
229
- "peak_cuda_reserved_bytes": {
230
- "values": [
231
- 3519021056.0,
232
- 3519021056.0,
233
- 3510632448.0
234
- ],
235
- "mean": 3516224853.3333335,
236
- "sample_sd": 4843165.086926248,
237
- "n": 3
238
- },
239
- "total_wall_seconds": {
240
- "values": [
241
- 1575.5905925639963,
242
- 1623.5639530190092,
243
- 1629.195114912989
244
- ],
245
- "mean": 1609.4498868319981,
246
- "sample_sd": 29.457874198835384,
247
- "n": 3
248
- },
249
- "test_seconds_per_article": {
250
- "values": [
251
- 0.07877608058201573,
252
- 0.08307423260041767,
253
- 0.08163602005436339
254
- ],
255
- "mean": 0.08116211107893226,
256
- "sample_sd": 0.002187914527825257,
257
- "n": 3
258
- }
259
- },
260
- "thresholds": [
261
- 0.05,
262
- 0.2,
263
- 0.2
264
- ]
265
- },
266
- "existing_specialized_separate_lineage": {
267
- "display": "Existing specialized 4L, separate lineage",
268
- "lineage": "historical task-specific shrink/distillation lineage",
269
- "metrics": {
270
- "macro_f1": {
271
- "values": [
272
- 0.5832848452288824,
273
- 0.5958405181160762,
274
- 0.5835785045218137
275
- ],
276
- "mean": 0.5875679559555907,
277
- "sample_sd": 0.007165753447433348,
278
- "n": 3
279
- },
280
- "micro_f1": {
281
- "values": [
282
- 0.6494252873563218,
283
- 0.6458616010854816,
284
- 0.6486486486486486
285
- ],
286
- "mean": 0.647978512363484,
287
- "sample_sd": 0.0018739735696546695,
288
- "n": 3
289
- },
290
- "exact_match": {
291
- "values": [
292
- 0.0,
293
- 0.0,
294
- 0.01818181818181818
295
- ],
296
- "mean": 0.006060606060606061,
297
- "sample_sd": 0.01049727762162956,
298
- "n": 3
299
- },
300
- "predicted_positive_rate": {
301
- "values": [
302
- 0.5597402597402598,
303
- 0.612987012987013,
304
- 0.5688311688311688
305
- ],
306
- "mean": 0.5805194805194805,
307
- "sample_sd": 0.02848274311618351,
308
- "n": 3
309
- }
310
- },
311
- "resources": {
312
- "parameters": {
313
- "values": [
314
- 210903360.0,
315
- 210903360.0,
316
- 210903360.0
317
- ],
318
- "mean": 210903360.0,
319
- "sample_sd": 0.0,
320
- "n": 3
321
- },
322
- "peak_cuda_allocated_bytes": {
323
- "values": [
324
- 2143655936.0,
325
- 2143655936.0,
326
- 2143655936.0
327
- ],
328
- "mean": 2143655936.0,
329
- "sample_sd": 0.0,
330
- "n": 3
331
- },
332
- "peak_cuda_reserved_bytes": {
333
- "values": [
334
- 2281701376.0,
335
- 2281701376.0,
336
- 2273312768.0
337
- ],
338
- "mean": 2278905173.3333335,
339
- "sample_sd": 4843165.086926248,
340
- "n": 3
341
- },
342
- "total_wall_seconds": {
343
- "values": [
344
- 1634.488557530014,
345
- 1652.9397718419787,
346
- 1643.488202479988
347
- ],
348
- "mean": 1643.6388439506602,
349
- "sample_sd": 9.226529522850198,
350
- "n": 3
351
- },
352
- "test_seconds_per_article": {
353
- "values": [
354
- 0.0847959352360869,
355
- 0.08598086021798239,
356
- 0.08552987641781908
357
- ],
358
- "mean": 0.08543555729062946,
359
- "sample_sd": 0.0005980667826383295,
360
- "n": 3
361
- }
362
- },
363
- "thresholds": [
364
- 0.1,
365
- 0.05,
366
- 0.05
367
- ]
368
- }
369
- },
370
- "same_seed_macro_f1_deltas_base_minus_control": [
371
- 0.07136687142547793,
372
- 0.06789732566779205,
373
- 0.06890378334579494
374
- ],
375
- "mean_macro_f1_delta_base_minus_control": 0.06938932681302164,
376
- "publication_gate_passed": true,
377
- "per_label_summary": [
378
- {
379
- "arm": "task_agnostic_base",
380
- "label": "Appeal_to_Authority",
381
- "support": 11,
382
- "precision_mean": 0.3958614484930274,
383
- "precision_sample_sd": 0.1238064457010119,
384
- "recall_mean": 0.5757575757575757,
385
- "recall_sample_sd": 0.05248638810814781,
386
- "f1_mean": 0.4657088122605364,
387
- "f1_sample_sd": 0.1020990126979069
388
- },
389
- {
390
- "arm": "task_agnostic_base",
391
- "label": "Appeal_to_fear-prejudice",
392
- "support": 24,
393
- "precision_mean": 0.5540244550770866,
394
- "precision_sample_sd": 0.02525673990402998,
395
- "recall_mean": 0.7361111111111112,
396
- "recall_sample_sd": 0.13393959390267993,
397
- "f1_mean": 0.6277521761392729,
398
- "f1_sample_sd": 0.04994879016183189
399
- },
400
- {
401
- "arm": "task_agnostic_base",
402
- "label": "Bandwagon,Reductio_ad_hitlerum",
403
- "support": 7,
404
- "precision_mean": 0.4851851851851852,
405
- "precision_sample_sd": 0.07882887649552875,
406
- "recall_mean": 0.8095238095238095,
407
- "recall_sample_sd": 0.08247860988423221,
408
- "f1_mean": 0.6006778309409889,
409
- "f1_sample_sd": 0.04793776258918466
410
- },
411
- {
412
- "arm": "task_agnostic_base",
413
- "label": "Black-and-White_Fallacy",
414
- "support": 12,
415
- "precision_mean": 0.3667929292929293,
416
- "precision_sample_sd": 0.07670649392711647,
417
- "recall_mean": 0.47222222222222227,
418
- "recall_sample_sd": 0.09622504486493764,
419
- "f1_mean": 0.4053892966936445,
420
- "f1_sample_sd": 0.04211369218048746
421
- },
422
- {
423
- "arm": "task_agnostic_base",
424
- "label": "Causal_Oversimplification",
425
- "support": 18,
426
- "precision_mean": 0.5428824049513704,
427
- "precision_sample_sd": 0.039216350192122156,
428
- "recall_mean": 0.7962962962962963,
429
- "recall_sample_sd": 0.11564811108145183,
430
- "f1_mean": 0.6446992723588468,
431
- "f1_sample_sd": 0.0634561308339175
432
- },
433
- {
434
- "arm": "task_agnostic_base",
435
- "label": "Doubt",
436
- "support": 22,
437
- "precision_mean": 0.5223832528180354,
438
- "precision_sample_sd": 0.06742780050556177,
439
- "recall_mean": 0.9848484848484849,
440
- "recall_sample_sd": 0.026243194054073875,
441
- "f1_mean": 0.6802057822343391,
442
- "f1_sample_sd": 0.04928561754686331
443
- },
444
- {
445
- "arm": "task_agnostic_base",
446
- "label": "Exaggeration,Minimisation",
447
- "support": 22,
448
- "precision_mean": 0.4620811287477954,
449
- "precision_sample_sd": 0.016801198309751,
450
- "recall_mean": 0.7727272727272727,
451
- "recall_sample_sd": 0.15745916432444337,
452
- "f1_mean": 0.5727040816326531,
453
- "f1_sample_sd": 0.03645259990419198
454
- },
455
- {
456
- "arm": "task_agnostic_base",
457
- "label": "Flag-Waving",
458
- "support": 17,
459
- "precision_mean": 0.5642642642642642,
460
- "precision_sample_sd": 0.1478145677656334,
461
- "recall_mean": 0.8235294117647058,
462
- "recall_sample_sd": 0.1764705882352941,
463
- "f1_mean": 0.6464506172839506,
464
- "f1_sample_sd": 0.035742219952758225
465
- },
466
- {
467
- "arm": "task_agnostic_base",
468
- "label": "Loaded_Language",
469
- "support": 45,
470
- "precision_mean": 0.8649032907793016,
471
- "precision_sample_sd": 0.046984312428427724,
472
- "recall_mean": 0.9851851851851852,
473
- "recall_sample_sd": 0.012830005981991702,
474
- "f1_mean": 0.9208741623431562,
475
- "f1_sample_sd": 0.03200926878737141
476
- },
477
- {
478
- "arm": "task_agnostic_base",
479
- "label": "Name_Calling,Labeling",
480
- "support": 33,
481
- "precision_mean": 0.7380952380952381,
482
- "precision_sample_sd": 0.0047619047619048005,
483
- "recall_mean": 0.9090909090909091,
484
- "recall_sample_sd": 0.10925912955951485,
485
- "f1_mean": 0.8125087983911513,
486
- "f1_sample_sd": 0.04252971464057227
487
- },
488
- {
489
- "arm": "task_agnostic_base",
490
- "label": "Repetition",
491
- "support": 23,
492
- "precision_mean": 0.5196620583717357,
493
- "precision_sample_sd": 0.021645909050194662,
494
- "recall_mean": 0.7536231884057971,
495
- "recall_sample_sd": 0.06641414050660639,
496
- "f1_mean": 0.6147520781519571,
497
- "f1_sample_sd": 0.03505927109416221
498
- },
499
- {
500
- "arm": "task_agnostic_base",
501
- "label": "Slogans",
502
- "support": 14,
503
- "precision_mean": 0.5161361626878869,
504
- "precision_sample_sd": 0.20305677930281646,
505
- "recall_mean": 0.7380952380952381,
506
- "recall_sample_sd": 0.10910894511799614,
507
- "f1_mean": 0.5834824090638044,
508
- "f1_sample_sd": 0.09862686289220693
509
- },
510
- {
511
- "arm": "task_agnostic_base",
512
- "label": "Thought-terminating_Cliches",
513
- "support": 7,
514
- "precision_mean": 0.3492063492063492,
515
- "precision_sample_sd": 0.143516994603374,
516
- "recall_mean": 0.3333333333333333,
517
- "recall_sample_sd": 0.08247860988423225,
518
- "f1_mean": 0.319014319014319,
519
- "f1_sample_sd": 0.04017590711117064
520
- },
521
- {
522
- "arm": "task_agnostic_base",
523
- "label": "Whataboutism,Straw_Men,Red_Herring",
524
- "support": 10,
525
- "precision_mean": 0.2761904761904762,
526
- "precision_sample_sd": 0.05302632726504782,
527
- "recall_mean": 0.5,
528
- "recall_sample_sd": 0.2645751311064591,
529
- "f1_mean": 0.330448343079922,
530
- "f1_sample_sd": 0.021845164720973014
531
- },
532
- {
533
- "arm": "structural_copy_control",
534
- "label": "Appeal_to_Authority",
535
- "support": 11,
536
- "precision_mean": 0.26003086419753085,
537
- "precision_sample_sd": 0.06281357095350094,
538
- "recall_mean": 0.5757575757575758,
539
- "recall_sample_sd": 0.262431940540739,
540
- "f1_mean": 0.35477582846003897,
541
- "f1_sample_sd": 0.11479479036518875
542
- },
543
- {
544
- "arm": "structural_copy_control",
545
- "label": "Appeal_to_fear-prejudice",
546
- "support": 24,
547
- "precision_mean": 0.5104813315339631,
548
- "precision_sample_sd": 0.038427508996734555,
549
- "recall_mean": 0.8611111111111112,
550
- "recall_sample_sd": 0.024056261216234387,
551
- "f1_mean": 0.6401745363877202,
552
- "f1_sample_sd": 0.029305467347150463
553
- },
554
- {
555
- "arm": "structural_copy_control",
556
- "label": "Bandwagon,Reductio_ad_hitlerum",
557
- "support": 7,
558
- "precision_mean": 0.2876623376623376,
559
- "precision_sample_sd": 0.061386823077618426,
560
- "recall_mean": 0.8571428571428571,
561
- "recall_sample_sd": 0.14285714285714285,
562
- "f1_mean": 0.4306391777656145,
563
- "f1_sample_sd": 0.08686392624022733
564
- },
565
- {
566
- "arm": "structural_copy_control",
567
- "label": "Black-and-White_Fallacy",
568
- "support": 12,
569
- "precision_mean": 0.31339031339031337,
570
- "precision_sample_sd": 0.08301311842925053,
571
- "recall_mean": 0.5277777777777778,
572
- "recall_sample_sd": 0.2678791878053599,
573
- "f1_mean": 0.38778427550357375,
574
- "f1_sample_sd": 0.13069755090515112
575
- },
576
- {
577
- "arm": "structural_copy_control",
578
- "label": "Causal_Oversimplification",
579
- "support": 18,
580
- "precision_mean": 0.4464201043148412,
581
- "precision_sample_sd": 0.013538446216471536,
582
- "recall_mean": 0.9259259259259259,
583
- "recall_sample_sd": 0.03207501495497923,
584
- "f1_mean": 0.6023809523809524,
585
- "f1_sample_sd": 0.018643640071557464
586
- },
587
- {
588
- "arm": "structural_copy_control",
589
- "label": "Doubt",
590
- "support": 22,
591
- "precision_mean": 0.4455314009661836,
592
- "precision_sample_sd": 0.04067072694857813,
593
- "recall_mean": 1.0,
594
- "recall_sample_sd": 0.0,
595
- "f1_mean": 0.615686274509804,
596
- "f1_sample_sd": 0.039427200694482675
597
- },
598
- {
599
- "arm": "structural_copy_control",
600
- "label": "Exaggeration,Minimisation",
601
- "support": 22,
602
- "precision_mean": 0.4364176150420584,
603
- "precision_sample_sd": 0.025093958106097894,
604
- "recall_mean": 0.8787878787878788,
605
- "recall_sample_sd": 0.05248638810814775,
606
- "f1_mean": 0.5829803047194351,
607
- "f1_sample_sd": 0.030899288356345502
608
- },
609
- {
610
- "arm": "structural_copy_control",
611
- "label": "Flag-Waving",
612
- "support": 17,
613
- "precision_mean": 0.3814814814814815,
614
- "precision_sample_sd": 0.05010277503136551,
615
- "recall_mean": 0.803921568627451,
616
- "recall_sample_sd": 0.17970885078258195,
617
- "f1_mean": 0.5138595287119652,
618
- "f1_sample_sd": 0.06400931989326974
619
- },
620
- {
621
- "arm": "structural_copy_control",
622
- "label": "Loaded_Language",
623
- "support": 45,
624
- "precision_mean": 0.8492581841638446,
625
- "precision_sample_sd": 0.016026591848331686,
626
- "recall_mean": 1.0,
627
- "recall_sample_sd": 0.0,
628
- "f1_mean": 0.9184311025253588,
629
- "f1_sample_sd": 0.009372233868132662
630
- },
631
- {
632
- "arm": "structural_copy_control",
633
- "label": "Name_Calling,Labeling",
634
- "support": 33,
635
- "precision_mean": 0.6375,
636
- "precision_sample_sd": 0.03410563654946854,
637
- "recall_mean": 0.9696969696969697,
638
- "recall_sample_sd": 0.030303030303030276,
639
- "f1_mean": 0.7685185185185185,
640
- "f1_sample_sd": 0.02023900779105559
641
- },
642
- {
643
- "arm": "structural_copy_control",
644
- "label": "Repetition",
645
- "support": 23,
646
- "precision_mean": 0.44863933452168747,
647
- "precision_sample_sd": 0.01520014564765896,
648
- "recall_mean": 0.9420289855072463,
649
- "recall_sample_sd": 0.06641414050660639,
650
- "f1_mean": 0.607248835423033,
651
- "f1_sample_sd": 0.019858627551810427
652
- },
653
- {
654
- "arm": "structural_copy_control",
655
- "label": "Slogans",
656
- "support": 14,
657
- "precision_mean": 0.3274318274318274,
658
- "precision_sample_sd": 0.05434598572773684,
659
- "recall_mean": 0.4523809523809524,
660
- "recall_sample_sd": 0.1486904285332952,
661
- "f1_mean": 0.3715828924162258,
662
- "f1_sample_sd": 0.07106660236343577
663
- },
664
- {
665
- "arm": "structural_copy_control",
666
- "label": "Thought-terminating_Cliches",
667
- "support": 7,
668
- "precision_mean": 0.1142156862745098,
669
- "precision_sample_sd": 0.012848374923608973,
670
- "recall_mean": 0.38095238095238093,
671
- "recall_sample_sd": 0.1649572197684645,
672
- "f1_mean": 0.17331433998100665,
673
- "f1_sample_sd": 0.029065880233243847
674
- },
675
- {
676
- "arm": "structural_copy_control",
677
- "label": "Whataboutism,Straw_Men,Red_Herring",
678
- "support": 10,
679
- "precision_mean": 0.19344598054275472,
680
- "precision_sample_sd": 0.03373872471827226,
681
- "recall_mean": 0.5666666666666667,
682
- "recall_sample_sd": 0.2081665999466133,
683
- "f1_mean": 0.28584083690299267,
684
- "f1_sample_sd": 0.06078854830133676
685
- },
686
- {
687
- "arm": "existing_specialized_separate_lineage",
688
- "label": "Appeal_to_Authority",
689
- "support": 11,
690
- "precision_mean": 0.37836021505376344,
691
- "precision_sample_sd": 0.10546365636105155,
692
- "recall_mean": 0.6666666666666666,
693
- "recall_sample_sd": 0.2287828616748712,
694
- "f1_mean": 0.46885361552028215,
695
- "f1_sample_sd": 0.09502747629890008
696
- },
697
- {
698
- "arm": "existing_specialized_separate_lineage",
699
- "label": "Appeal_to_fear-prejudice",
700
- "support": 24,
701
- "precision_mean": 0.525708635464733,
702
- "precision_sample_sd": 0.02235198659615819,
703
- "recall_mean": 0.8888888888888888,
704
- "recall_sample_sd": 0.048112522432468774,
705
- "f1_mean": 0.6599065351235072,
706
- "f1_sample_sd": 0.015362381542320084
707
- },
708
- {
709
- "arm": "existing_specialized_separate_lineage",
710
- "label": "Bandwagon,Reductio_ad_hitlerum",
711
- "support": 7,
712
- "precision_mean": 0.44485144485144484,
713
- "precision_sample_sd": 0.016487284575115283,
714
- "recall_mean": 0.7619047619047619,
715
- "recall_sample_sd": 0.1649572197684645,
716
- "f1_mean": 0.5571428571428572,
717
- "f1_sample_sd": 0.051507875363771265
718
- },
719
- {
720
- "arm": "existing_specialized_separate_lineage",
721
- "label": "Black-and-White_Fallacy",
722
- "support": 12,
723
- "precision_mean": 0.4662644544997486,
724
- "precision_sample_sd": 0.004538274146363688,
725
- "recall_mean": 0.5833333333333334,
726
- "recall_sample_sd": 0.08333333333333331,
727
- "f1_mean": 0.516747552149851,
728
- "f1_sample_sd": 0.03589484965697345
729
- },
730
- {
731
- "arm": "existing_specialized_separate_lineage",
732
- "label": "Causal_Oversimplification",
733
- "support": 18,
734
- "precision_mean": 0.4472934472934473,
735
- "precision_sample_sd": 0.013055771210700413,
736
- "recall_mean": 0.8333333333333333,
737
- "recall_sample_sd": 0.1469861839480328,
738
- "f1_mean": 0.5781794553724379,
739
- "f1_sample_sd": 0.028407563015813122
740
- },
741
- {
742
- "arm": "existing_specialized_separate_lineage",
743
- "label": "Doubt",
744
- "support": 22,
745
- "precision_mean": 0.5045661001171619,
746
- "precision_sample_sd": 0.023581174969843577,
747
- "recall_mean": 1.0,
748
- "recall_sample_sd": 0.0,
749
- "f1_mean": 0.6704939668174963,
750
- "f1_sample_sd": 0.020973132363659372
751
- },
752
- {
753
- "arm": "existing_specialized_separate_lineage",
754
- "label": "Exaggeration,Minimisation",
755
- "support": 22,
756
- "precision_mean": 0.44589793281653745,
757
- "precision_sample_sd": 0.010987847993006481,
758
- "recall_mean": 0.9393939393939394,
759
- "recall_sample_sd": 0.06943296507508849,
760
- "f1_mean": 0.6043956043956044,
761
- "f1_sample_sd": 0.022305256187015846
762
- },
763
- {
764
- "arm": "existing_specialized_separate_lineage",
765
- "label": "Flag-Waving",
766
- "support": 17,
767
- "precision_mean": 0.5576388888888889,
768
- "precision_sample_sd": 0.03705367103320583,
769
- "recall_mean": 0.8823529411764706,
770
- "recall_sample_sd": 0.11764705882352944,
771
- "f1_mean": 0.680769868923179,
772
- "f1_sample_sd": 0.04164658769406446
773
- },
774
- {
775
- "arm": "existing_specialized_separate_lineage",
776
- "label": "Loaded_Language",
777
- "support": 45,
778
- "precision_mean": 0.8387650085763294,
779
- "precision_sample_sd": 0.01782556577257849,
780
- "recall_mean": 1.0,
781
- "recall_sample_sd": 0.0,
782
- "f1_mean": 0.9122448979591837,
783
- "f1_sample_sd": 0.01060439269940129
784
- },
785
- {
786
- "arm": "existing_specialized_separate_lineage",
787
- "label": "Name_Calling,Labeling",
788
- "support": 33,
789
- "precision_mean": 0.6957104190221673,
790
- "precision_sample_sd": 0.019201490809269257,
791
- "recall_mean": 0.9696969696969697,
792
- "recall_sample_sd": 0.030303030303030276,
793
- "f1_mean": 0.8101049442821594,
794
- "f1_sample_sd": 0.022092733328802537
795
- },
796
- {
797
- "arm": "existing_specialized_separate_lineage",
798
- "label": "Repetition",
799
- "support": 23,
800
- "precision_mean": 0.4781557067271353,
801
- "precision_sample_sd": 0.009897987322673213,
802
- "recall_mean": 0.9420289855072463,
803
- "recall_sample_sd": 0.06641414050660639,
804
- "f1_mean": 0.6337774426009719,
805
- "f1_sample_sd": 0.016444118333485254
806
- },
807
- {
808
- "arm": "existing_specialized_separate_lineage",
809
- "label": "Slogans",
810
- "support": 14,
811
- "precision_mean": 0.4334215167548501,
812
- "precision_sample_sd": 0.05811581861282867,
813
- "recall_mean": 0.6904761904761905,
814
- "recall_sample_sd": 0.1649572197684645,
815
- "f1_mean": 0.5201316298877274,
816
- "f1_sample_sd": 0.018567913895656114
817
- },
818
- {
819
- "arm": "existing_specialized_separate_lineage",
820
- "label": "Thought-terminating_Cliches",
821
- "support": 7,
822
- "precision_mean": 0.18446115288220552,
823
- "precision_sample_sd": 0.03641251252864963,
824
- "recall_mean": 0.6666666666666666,
825
- "recall_sample_sd": 0.1649572197684645,
826
- "f1_mean": 0.2868626868626869,
827
- "f1_sample_sd": 0.0511620864429468
828
- },
829
- {
830
- "arm": "existing_specialized_separate_lineage",
831
- "label": "Whataboutism,Straw_Men,Red_Herring",
832
- "support": 10,
833
- "precision_mean": 0.2777777777777778,
834
- "precision_sample_sd": 0.0481125224324688,
835
- "recall_mean": 0.4,
836
- "recall_sample_sd": 0.0,
837
- "f1_mean": 0.32634032634032634,
838
- "f1_sample_sd": 0.03229931575886014
839
- }
840
- ],
841
- "compression_ladder": [
842
- {
843
- "name": "24L upstream text teacher",
844
- "layers": 24,
845
- "parameters": 752393024,
846
- "weight_bytes": 1504825632,
847
- "weight_mib": 1435.1135559082031,
848
- "weight_sha256": "30783da4651259c26b8a204821ea426b9526431e0c14da0d1c6fa284e7414e5e",
849
- "kd_stage": "",
850
- "mean_total_loss": "",
851
- "mean_interface_loss": "",
852
- "mean_final_loss": ""
853
- },
854
- {
855
- "name": "8L task-free KD",
856
- "layers": 8,
857
- "parameters": 420318144,
858
- "weight_bytes": 840647880,
859
- "weight_mib": 801.7042922973633,
860
- "weight_sha256": "086fbaf9a4838ecde56cd3541c9c3212b22d0c5ee745cf2c1c5e0bb31e0ec22c",
861
- "kd_stage": "24to8",
862
- "mean_total_loss": 0.17622653172702485,
863
- "mean_interface_loss": 0.21562811763578793,
864
- "mean_final_loss": 0.16626427527853593
865
- },
866
- {
867
- "name": "6L task-free KD",
868
- "layers": 6,
869
- "parameters": 377207424,
870
- "weight_bytes": 754423424,
871
- "weight_mib": 719.4742431640625,
872
- "weight_sha256": "492749c20f77a42f6bc993f0927a82217b94f872b7eacf25612f7520c9bc28ce",
873
- "kd_stage": "8to6",
874
- "mean_total_loss": 0.03382104352249371,
875
- "mean_interface_loss": 0.04486069045515251,
876
- "mean_final_loss": 0.031147060785087888
877
- },
878
- {
879
- "name": "4L task-free KD root",
880
- "layers": 4,
881
- "parameters": 334096704,
882
- "weight_bytes": 668198976,
883
- "weight_mib": 637.2442016601562,
884
- "weight_sha256": "2732c616772fe320cdea228ab4554981418b1b2bf615c4183fb1ac8e6e2168d3",
885
- "kd_stage": "6to4",
886
- "mean_total_loss": 0.047020394468859195,
887
- "mean_interface_loss": 0.07048421185527332,
888
- "mean_final_loss": 0.043119939065377366
889
- }
890
- ]
891
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
benchmark/summary.json DELETED
@@ -1,891 +0,0 @@
1
- {
2
- "schema": "standalone4l-complete-transfer-benchmark-v2",
3
- "status": "complete_from_frozen_reports",
4
- "claim_scope": "Exploratory fresh-head transfer evidence on one English task and a previously opened 55-article SemEval-derived test split.",
5
- "headless_root_requires_adaptation": true,
6
- "test_previously_opened": true,
7
- "test_articles": 55,
8
- "test_windows": 434,
9
- "seeds": [
10
- 41,
11
- 42,
12
- 43
13
- ],
14
- "three_seed_sd_is_not_confidence_interval": true,
15
- "multilingual_downstream_evaluation_performed": false,
16
- "upstream_multilingual_metadata": {
17
- "source": "Qwen/Qwen3.5-0.8B official model card",
18
- "languages_and_dialects_claimed_upstream": 201,
19
- "scope": "inherited tokenizer and architecture coverage only"
20
- },
21
- "protocol": {
22
- "epochs": 5,
23
- "primary_epoch": 5,
24
- "early_stopping": false,
25
- "max_length": 256,
26
- "stride": 128,
27
- "window_target": "span-aligned technique labels",
28
- "article_aggregation": "labelwise maximum probability",
29
- "micro_batch_size": 1,
30
- "gradient_accumulation": 32,
31
- "effective_batch_size": 32,
32
- "learning_rate": 2e-05,
33
- "weight_decay": 0.01,
34
- "gradient_clip": 1.0,
35
- "optimizer": "AdamW",
36
- "dtype": "bfloat16",
37
- "thresholds": [
38
- 0.05,
39
- 0.1,
40
- 0.15,
41
- 0.2,
42
- 0.25,
43
- 0.3,
44
- 0.35,
45
- 0.4,
46
- 0.45,
47
- 0.5,
48
- 0.55,
49
- 0.6,
50
- 0.65,
51
- 0.7,
52
- 0.75,
53
- 0.8,
54
- 0.85,
55
- 0.9
56
- ],
57
- "threshold_selection": "epoch-5 calibration macro-F1, then micro-F1"
58
- },
59
- "arms": {
60
- "task_agnostic_base": {
61
- "display": "Task-agnostic KD base",
62
- "lineage": "root base plus a fresh 14-label transfer head",
63
- "metrics": {
64
- "macro_f1": {
65
- "values": [
66
- 0.5883093293517536,
67
- 0.6028379036218096,
68
- 0.5712816197954101
69
- ],
70
- "mean": 0.5874762842563245,
71
- "sample_sd": 0.01579462678670608,
72
- "n": 3
73
- },
74
- "micro_f1": {
75
- "values": [
76
- 0.6468085106382978,
77
- 0.6655172413793103,
78
- 0.6465256797583082
79
- ],
80
- "mean": 0.6529504772586388,
81
- "sample_sd": 0.010884055708931178,
82
- "n": 3
83
- },
84
- "exact_match": {
85
- "values": [
86
- 0.05454545454545454,
87
- 0.03636363636363636,
88
- 0.0
89
- ],
90
- "mean": 0.0303030303030303,
91
- "sample_sd": 0.027773186030035393,
92
- "n": 3
93
- },
94
- "predicted_positive_rate": {
95
- "values": [
96
- 0.5714285714285714,
97
- 0.4090909090909091,
98
- 0.5155844155844156
99
- ],
100
- "mean": 0.4987012987012987,
101
- "sample_sd": 0.08247520160679181,
102
- "n": 3
103
- }
104
- },
105
- "resources": {
106
- "parameters": {
107
- "values": [
108
- 334111040.0,
109
- 334111040.0,
110
- 334111040.0
111
- ],
112
- "mean": 334111040.0,
113
- "sample_sd": 0.0,
114
- "n": 3
115
- },
116
- "peak_cuda_allocated_bytes": {
117
- "values": [
118
- 3380975616.0,
119
- 3380975616.0,
120
- 3380975616.0
121
- ],
122
- "mean": 3380975616.0,
123
- "sample_sd": 0.0,
124
- "n": 3
125
- },
126
- "peak_cuda_reserved_bytes": {
127
- "values": [
128
- 3519021056.0,
129
- 3519021056.0,
130
- 3510632448.0
131
- ],
132
- "mean": 3516224853.3333335,
133
- "sample_sd": 4843165.086926248,
134
- "n": 3
135
- },
136
- "total_wall_seconds": {
137
- "values": [
138
- 1583.0297977919981,
139
- 1588.617994428001,
140
- 1603.4310500120046
141
- ],
142
- "mean": 1591.692947410668,
143
- "sample_sd": 10.54249851412311,
144
- "n": 3
145
- },
146
- "test_seconds_per_article": {
147
- "values": [
148
- 0.08164824154545468,
149
- 0.07871871667289683,
150
- 0.08355205721825107
151
- ],
152
- "mean": 0.08130633847886752,
153
- "sample_sd": 0.0024347419754560395,
154
- "n": 3
155
- }
156
- },
157
- "thresholds": [
158
- 0.15,
159
- 0.3,
160
- 0.1
161
- ]
162
- },
163
- "structural_copy_control": {
164
- "display": "Structural copy, no task-free KD",
165
- "lineage": "same four-layer structure without task-free KD",
166
- "metrics": {
167
- "macro_f1": {
168
- "values": [
169
- 0.5169424579262757,
170
- 0.5349405779540175,
171
- 0.5023778364496152
172
- ],
173
- "mean": 0.5180869574433028,
174
- "sample_sd": 0.016311512588628545,
175
- "n": 3
176
- },
177
- "micro_f1": {
178
- "values": [
179
- 0.5773955773955775,
180
- 0.6083445491251682,
181
- 0.585635359116022
182
- ],
183
- "mean": 0.5904584952122559,
184
- "sample_sd": 0.01602830914762526,
185
- "n": 3
186
- },
187
- "exact_match": {
188
- "values": [
189
- 0.0,
190
- 0.0,
191
- 0.0
192
- ],
193
- "mean": 0.0,
194
- "sample_sd": 0.0,
195
- "n": 3
196
- },
197
- "predicted_positive_rate": {
198
- "values": [
199
- 0.712987012987013,
200
- 0.6207792207792208,
201
- 0.5961038961038961
202
- ],
203
- "mean": 0.6432900432900432,
204
- "sample_sd": 0.06160737456395461,
205
- "n": 3
206
- }
207
- },
208
- "resources": {
209
- "parameters": {
210
- "values": [
211
- 334111040.0,
212
- 334111040.0,
213
- 334111040.0
214
- ],
215
- "mean": 334111040.0,
216
- "sample_sd": 0.0,
217
- "n": 3
218
- },
219
- "peak_cuda_allocated_bytes": {
220
- "values": [
221
- 3380975616.0,
222
- 3380975616.0,
223
- 3380975616.0
224
- ],
225
- "mean": 3380975616.0,
226
- "sample_sd": 0.0,
227
- "n": 3
228
- },
229
- "peak_cuda_reserved_bytes": {
230
- "values": [
231
- 3519021056.0,
232
- 3519021056.0,
233
- 3510632448.0
234
- ],
235
- "mean": 3516224853.3333335,
236
- "sample_sd": 4843165.086926248,
237
- "n": 3
238
- },
239
- "total_wall_seconds": {
240
- "values": [
241
- 1575.5905925639963,
242
- 1623.5639530190092,
243
- 1629.195114912989
244
- ],
245
- "mean": 1609.4498868319981,
246
- "sample_sd": 29.457874198835384,
247
- "n": 3
248
- },
249
- "test_seconds_per_article": {
250
- "values": [
251
- 0.07877608058201573,
252
- 0.08307423260041767,
253
- 0.08163602005436339
254
- ],
255
- "mean": 0.08116211107893226,
256
- "sample_sd": 0.002187914527825257,
257
- "n": 3
258
- }
259
- },
260
- "thresholds": [
261
- 0.05,
262
- 0.2,
263
- 0.2
264
- ]
265
- },
266
- "existing_specialized_separate_lineage": {
267
- "display": "Existing specialized 4L, separate lineage",
268
- "lineage": "historical task-specific shrink/distillation lineage",
269
- "metrics": {
270
- "macro_f1": {
271
- "values": [
272
- 0.5832848452288824,
273
- 0.5958405181160762,
274
- 0.5835785045218137
275
- ],
276
- "mean": 0.5875679559555907,
277
- "sample_sd": 0.007165753447433348,
278
- "n": 3
279
- },
280
- "micro_f1": {
281
- "values": [
282
- 0.6494252873563218,
283
- 0.6458616010854816,
284
- 0.6486486486486486
285
- ],
286
- "mean": 0.647978512363484,
287
- "sample_sd": 0.0018739735696546695,
288
- "n": 3
289
- },
290
- "exact_match": {
291
- "values": [
292
- 0.0,
293
- 0.0,
294
- 0.01818181818181818
295
- ],
296
- "mean": 0.006060606060606061,
297
- "sample_sd": 0.01049727762162956,
298
- "n": 3
299
- },
300
- "predicted_positive_rate": {
301
- "values": [
302
- 0.5597402597402598,
303
- 0.612987012987013,
304
- 0.5688311688311688
305
- ],
306
- "mean": 0.5805194805194805,
307
- "sample_sd": 0.02848274311618351,
308
- "n": 3
309
- }
310
- },
311
- "resources": {
312
- "parameters": {
313
- "values": [
314
- 210903360.0,
315
- 210903360.0,
316
- 210903360.0
317
- ],
318
- "mean": 210903360.0,
319
- "sample_sd": 0.0,
320
- "n": 3
321
- },
322
- "peak_cuda_allocated_bytes": {
323
- "values": [
324
- 2143655936.0,
325
- 2143655936.0,
326
- 2143655936.0
327
- ],
328
- "mean": 2143655936.0,
329
- "sample_sd": 0.0,
330
- "n": 3
331
- },
332
- "peak_cuda_reserved_bytes": {
333
- "values": [
334
- 2281701376.0,
335
- 2281701376.0,
336
- 2273312768.0
337
- ],
338
- "mean": 2278905173.3333335,
339
- "sample_sd": 4843165.086926248,
340
- "n": 3
341
- },
342
- "total_wall_seconds": {
343
- "values": [
344
- 1634.488557530014,
345
- 1652.9397718419787,
346
- 1643.488202479988
347
- ],
348
- "mean": 1643.6388439506602,
349
- "sample_sd": 9.226529522850198,
350
- "n": 3
351
- },
352
- "test_seconds_per_article": {
353
- "values": [
354
- 0.0847959352360869,
355
- 0.08598086021798239,
356
- 0.08552987641781908
357
- ],
358
- "mean": 0.08543555729062946,
359
- "sample_sd": 0.0005980667826383295,
360
- "n": 3
361
- }
362
- },
363
- "thresholds": [
364
- 0.1,
365
- 0.05,
366
- 0.05
367
- ]
368
- }
369
- },
370
- "same_seed_macro_f1_deltas_base_minus_control": [
371
- 0.07136687142547793,
372
- 0.06789732566779205,
373
- 0.06890378334579494
374
- ],
375
- "mean_macro_f1_delta_base_minus_control": 0.06938932681302164,
376
- "publication_gate_passed": true,
377
- "per_label_summary": [
378
- {
379
- "arm": "task_agnostic_base",
380
- "label": "Appeal_to_Authority",
381
- "support": 11,
382
- "precision_mean": 0.3958614484930274,
383
- "precision_sample_sd": 0.1238064457010119,
384
- "recall_mean": 0.5757575757575757,
385
- "recall_sample_sd": 0.05248638810814781,
386
- "f1_mean": 0.4657088122605364,
387
- "f1_sample_sd": 0.1020990126979069
388
- },
389
- {
390
- "arm": "task_agnostic_base",
391
- "label": "Appeal_to_fear-prejudice",
392
- "support": 24,
393
- "precision_mean": 0.5540244550770866,
394
- "precision_sample_sd": 0.02525673990402998,
395
- "recall_mean": 0.7361111111111112,
396
- "recall_sample_sd": 0.13393959390267993,
397
- "f1_mean": 0.6277521761392729,
398
- "f1_sample_sd": 0.04994879016183189
399
- },
400
- {
401
- "arm": "task_agnostic_base",
402
- "label": "Bandwagon,Reductio_ad_hitlerum",
403
- "support": 7,
404
- "precision_mean": 0.4851851851851852,
405
- "precision_sample_sd": 0.07882887649552875,
406
- "recall_mean": 0.8095238095238095,
407
- "recall_sample_sd": 0.08247860988423221,
408
- "f1_mean": 0.6006778309409889,
409
- "f1_sample_sd": 0.04793776258918466
410
- },
411
- {
412
- "arm": "task_agnostic_base",
413
- "label": "Black-and-White_Fallacy",
414
- "support": 12,
415
- "precision_mean": 0.3667929292929293,
416
- "precision_sample_sd": 0.07670649392711647,
417
- "recall_mean": 0.47222222222222227,
418
- "recall_sample_sd": 0.09622504486493764,
419
- "f1_mean": 0.4053892966936445,
420
- "f1_sample_sd": 0.04211369218048746
421
- },
422
- {
423
- "arm": "task_agnostic_base",
424
- "label": "Causal_Oversimplification",
425
- "support": 18,
426
- "precision_mean": 0.5428824049513704,
427
- "precision_sample_sd": 0.039216350192122156,
428
- "recall_mean": 0.7962962962962963,
429
- "recall_sample_sd": 0.11564811108145183,
430
- "f1_mean": 0.6446992723588468,
431
- "f1_sample_sd": 0.0634561308339175
432
- },
433
- {
434
- "arm": "task_agnostic_base",
435
- "label": "Doubt",
436
- "support": 22,
437
- "precision_mean": 0.5223832528180354,
438
- "precision_sample_sd": 0.06742780050556177,
439
- "recall_mean": 0.9848484848484849,
440
- "recall_sample_sd": 0.026243194054073875,
441
- "f1_mean": 0.6802057822343391,
442
- "f1_sample_sd": 0.04928561754686331
443
- },
444
- {
445
- "arm": "task_agnostic_base",
446
- "label": "Exaggeration,Minimisation",
447
- "support": 22,
448
- "precision_mean": 0.4620811287477954,
449
- "precision_sample_sd": 0.016801198309751,
450
- "recall_mean": 0.7727272727272727,
451
- "recall_sample_sd": 0.15745916432444337,
452
- "f1_mean": 0.5727040816326531,
453
- "f1_sample_sd": 0.03645259990419198
454
- },
455
- {
456
- "arm": "task_agnostic_base",
457
- "label": "Flag-Waving",
458
- "support": 17,
459
- "precision_mean": 0.5642642642642642,
460
- "precision_sample_sd": 0.1478145677656334,
461
- "recall_mean": 0.8235294117647058,
462
- "recall_sample_sd": 0.1764705882352941,
463
- "f1_mean": 0.6464506172839506,
464
- "f1_sample_sd": 0.035742219952758225
465
- },
466
- {
467
- "arm": "task_agnostic_base",
468
- "label": "Loaded_Language",
469
- "support": 45,
470
- "precision_mean": 0.8649032907793016,
471
- "precision_sample_sd": 0.046984312428427724,
472
- "recall_mean": 0.9851851851851852,
473
- "recall_sample_sd": 0.012830005981991702,
474
- "f1_mean": 0.9208741623431562,
475
- "f1_sample_sd": 0.03200926878737141
476
- },
477
- {
478
- "arm": "task_agnostic_base",
479
- "label": "Name_Calling,Labeling",
480
- "support": 33,
481
- "precision_mean": 0.7380952380952381,
482
- "precision_sample_sd": 0.0047619047619048005,
483
- "recall_mean": 0.9090909090909091,
484
- "recall_sample_sd": 0.10925912955951485,
485
- "f1_mean": 0.8125087983911513,
486
- "f1_sample_sd": 0.04252971464057227
487
- },
488
- {
489
- "arm": "task_agnostic_base",
490
- "label": "Repetition",
491
- "support": 23,
492
- "precision_mean": 0.5196620583717357,
493
- "precision_sample_sd": 0.021645909050194662,
494
- "recall_mean": 0.7536231884057971,
495
- "recall_sample_sd": 0.06641414050660639,
496
- "f1_mean": 0.6147520781519571,
497
- "f1_sample_sd": 0.03505927109416221
498
- },
499
- {
500
- "arm": "task_agnostic_base",
501
- "label": "Slogans",
502
- "support": 14,
503
- "precision_mean": 0.5161361626878869,
504
- "precision_sample_sd": 0.20305677930281646,
505
- "recall_mean": 0.7380952380952381,
506
- "recall_sample_sd": 0.10910894511799614,
507
- "f1_mean": 0.5834824090638044,
508
- "f1_sample_sd": 0.09862686289220693
509
- },
510
- {
511
- "arm": "task_agnostic_base",
512
- "label": "Thought-terminating_Cliches",
513
- "support": 7,
514
- "precision_mean": 0.3492063492063492,
515
- "precision_sample_sd": 0.143516994603374,
516
- "recall_mean": 0.3333333333333333,
517
- "recall_sample_sd": 0.08247860988423225,
518
- "f1_mean": 0.319014319014319,
519
- "f1_sample_sd": 0.04017590711117064
520
- },
521
- {
522
- "arm": "task_agnostic_base",
523
- "label": "Whataboutism,Straw_Men,Red_Herring",
524
- "support": 10,
525
- "precision_mean": 0.2761904761904762,
526
- "precision_sample_sd": 0.05302632726504782,
527
- "recall_mean": 0.5,
528
- "recall_sample_sd": 0.2645751311064591,
529
- "f1_mean": 0.330448343079922,
530
- "f1_sample_sd": 0.021845164720973014
531
- },
532
- {
533
- "arm": "structural_copy_control",
534
- "label": "Appeal_to_Authority",
535
- "support": 11,
536
- "precision_mean": 0.26003086419753085,
537
- "precision_sample_sd": 0.06281357095350094,
538
- "recall_mean": 0.5757575757575758,
539
- "recall_sample_sd": 0.262431940540739,
540
- "f1_mean": 0.35477582846003897,
541
- "f1_sample_sd": 0.11479479036518875
542
- },
543
- {
544
- "arm": "structural_copy_control",
545
- "label": "Appeal_to_fear-prejudice",
546
- "support": 24,
547
- "precision_mean": 0.5104813315339631,
548
- "precision_sample_sd": 0.038427508996734555,
549
- "recall_mean": 0.8611111111111112,
550
- "recall_sample_sd": 0.024056261216234387,
551
- "f1_mean": 0.6401745363877202,
552
- "f1_sample_sd": 0.029305467347150463
553
- },
554
- {
555
- "arm": "structural_copy_control",
556
- "label": "Bandwagon,Reductio_ad_hitlerum",
557
- "support": 7,
558
- "precision_mean": 0.2876623376623376,
559
- "precision_sample_sd": 0.061386823077618426,
560
- "recall_mean": 0.8571428571428571,
561
- "recall_sample_sd": 0.14285714285714285,
562
- "f1_mean": 0.4306391777656145,
563
- "f1_sample_sd": 0.08686392624022733
564
- },
565
- {
566
- "arm": "structural_copy_control",
567
- "label": "Black-and-White_Fallacy",
568
- "support": 12,
569
- "precision_mean": 0.31339031339031337,
570
- "precision_sample_sd": 0.08301311842925053,
571
- "recall_mean": 0.5277777777777778,
572
- "recall_sample_sd": 0.2678791878053599,
573
- "f1_mean": 0.38778427550357375,
574
- "f1_sample_sd": 0.13069755090515112
575
- },
576
- {
577
- "arm": "structural_copy_control",
578
- "label": "Causal_Oversimplification",
579
- "support": 18,
580
- "precision_mean": 0.4464201043148412,
581
- "precision_sample_sd": 0.013538446216471536,
582
- "recall_mean": 0.9259259259259259,
583
- "recall_sample_sd": 0.03207501495497923,
584
- "f1_mean": 0.6023809523809524,
585
- "f1_sample_sd": 0.018643640071557464
586
- },
587
- {
588
- "arm": "structural_copy_control",
589
- "label": "Doubt",
590
- "support": 22,
591
- "precision_mean": 0.4455314009661836,
592
- "precision_sample_sd": 0.04067072694857813,
593
- "recall_mean": 1.0,
594
- "recall_sample_sd": 0.0,
595
- "f1_mean": 0.615686274509804,
596
- "f1_sample_sd": 0.039427200694482675
597
- },
598
- {
599
- "arm": "structural_copy_control",
600
- "label": "Exaggeration,Minimisation",
601
- "support": 22,
602
- "precision_mean": 0.4364176150420584,
603
- "precision_sample_sd": 0.025093958106097894,
604
- "recall_mean": 0.8787878787878788,
605
- "recall_sample_sd": 0.05248638810814775,
606
- "f1_mean": 0.5829803047194351,
607
- "f1_sample_sd": 0.030899288356345502
608
- },
609
- {
610
- "arm": "structural_copy_control",
611
- "label": "Flag-Waving",
612
- "support": 17,
613
- "precision_mean": 0.3814814814814815,
614
- "precision_sample_sd": 0.05010277503136551,
615
- "recall_mean": 0.803921568627451,
616
- "recall_sample_sd": 0.17970885078258195,
617
- "f1_mean": 0.5138595287119652,
618
- "f1_sample_sd": 0.06400931989326974
619
- },
620
- {
621
- "arm": "structural_copy_control",
622
- "label": "Loaded_Language",
623
- "support": 45,
624
- "precision_mean": 0.8492581841638446,
625
- "precision_sample_sd": 0.016026591848331686,
626
- "recall_mean": 1.0,
627
- "recall_sample_sd": 0.0,
628
- "f1_mean": 0.9184311025253588,
629
- "f1_sample_sd": 0.009372233868132662
630
- },
631
- {
632
- "arm": "structural_copy_control",
633
- "label": "Name_Calling,Labeling",
634
- "support": 33,
635
- "precision_mean": 0.6375,
636
- "precision_sample_sd": 0.03410563654946854,
637
- "recall_mean": 0.9696969696969697,
638
- "recall_sample_sd": 0.030303030303030276,
639
- "f1_mean": 0.7685185185185185,
640
- "f1_sample_sd": 0.02023900779105559
641
- },
642
- {
643
- "arm": "structural_copy_control",
644
- "label": "Repetition",
645
- "support": 23,
646
- "precision_mean": 0.44863933452168747,
647
- "precision_sample_sd": 0.01520014564765896,
648
- "recall_mean": 0.9420289855072463,
649
- "recall_sample_sd": 0.06641414050660639,
650
- "f1_mean": 0.607248835423033,
651
- "f1_sample_sd": 0.019858627551810427
652
- },
653
- {
654
- "arm": "structural_copy_control",
655
- "label": "Slogans",
656
- "support": 14,
657
- "precision_mean": 0.3274318274318274,
658
- "precision_sample_sd": 0.05434598572773684,
659
- "recall_mean": 0.4523809523809524,
660
- "recall_sample_sd": 0.1486904285332952,
661
- "f1_mean": 0.3715828924162258,
662
- "f1_sample_sd": 0.07106660236343577
663
- },
664
- {
665
- "arm": "structural_copy_control",
666
- "label": "Thought-terminating_Cliches",
667
- "support": 7,
668
- "precision_mean": 0.1142156862745098,
669
- "precision_sample_sd": 0.012848374923608973,
670
- "recall_mean": 0.38095238095238093,
671
- "recall_sample_sd": 0.1649572197684645,
672
- "f1_mean": 0.17331433998100665,
673
- "f1_sample_sd": 0.029065880233243847
674
- },
675
- {
676
- "arm": "structural_copy_control",
677
- "label": "Whataboutism,Straw_Men,Red_Herring",
678
- "support": 10,
679
- "precision_mean": 0.19344598054275472,
680
- "precision_sample_sd": 0.03373872471827226,
681
- "recall_mean": 0.5666666666666667,
682
- "recall_sample_sd": 0.2081665999466133,
683
- "f1_mean": 0.28584083690299267,
684
- "f1_sample_sd": 0.06078854830133676
685
- },
686
- {
687
- "arm": "existing_specialized_separate_lineage",
688
- "label": "Appeal_to_Authority",
689
- "support": 11,
690
- "precision_mean": 0.37836021505376344,
691
- "precision_sample_sd": 0.10546365636105155,
692
- "recall_mean": 0.6666666666666666,
693
- "recall_sample_sd": 0.2287828616748712,
694
- "f1_mean": 0.46885361552028215,
695
- "f1_sample_sd": 0.09502747629890008
696
- },
697
- {
698
- "arm": "existing_specialized_separate_lineage",
699
- "label": "Appeal_to_fear-prejudice",
700
- "support": 24,
701
- "precision_mean": 0.525708635464733,
702
- "precision_sample_sd": 0.02235198659615819,
703
- "recall_mean": 0.8888888888888888,
704
- "recall_sample_sd": 0.048112522432468774,
705
- "f1_mean": 0.6599065351235072,
706
- "f1_sample_sd": 0.015362381542320084
707
- },
708
- {
709
- "arm": "existing_specialized_separate_lineage",
710
- "label": "Bandwagon,Reductio_ad_hitlerum",
711
- "support": 7,
712
- "precision_mean": 0.44485144485144484,
713
- "precision_sample_sd": 0.016487284575115283,
714
- "recall_mean": 0.7619047619047619,
715
- "recall_sample_sd": 0.1649572197684645,
716
- "f1_mean": 0.5571428571428572,
717
- "f1_sample_sd": 0.051507875363771265
718
- },
719
- {
720
- "arm": "existing_specialized_separate_lineage",
721
- "label": "Black-and-White_Fallacy",
722
- "support": 12,
723
- "precision_mean": 0.4662644544997486,
724
- "precision_sample_sd": 0.004538274146363688,
725
- "recall_mean": 0.5833333333333334,
726
- "recall_sample_sd": 0.08333333333333331,
727
- "f1_mean": 0.516747552149851,
728
- "f1_sample_sd": 0.03589484965697345
729
- },
730
- {
731
- "arm": "existing_specialized_separate_lineage",
732
- "label": "Causal_Oversimplification",
733
- "support": 18,
734
- "precision_mean": 0.4472934472934473,
735
- "precision_sample_sd": 0.013055771210700413,
736
- "recall_mean": 0.8333333333333333,
737
- "recall_sample_sd": 0.1469861839480328,
738
- "f1_mean": 0.5781794553724379,
739
- "f1_sample_sd": 0.028407563015813122
740
- },
741
- {
742
- "arm": "existing_specialized_separate_lineage",
743
- "label": "Doubt",
744
- "support": 22,
745
- "precision_mean": 0.5045661001171619,
746
- "precision_sample_sd": 0.023581174969843577,
747
- "recall_mean": 1.0,
748
- "recall_sample_sd": 0.0,
749
- "f1_mean": 0.6704939668174963,
750
- "f1_sample_sd": 0.020973132363659372
751
- },
752
- {
753
- "arm": "existing_specialized_separate_lineage",
754
- "label": "Exaggeration,Minimisation",
755
- "support": 22,
756
- "precision_mean": 0.44589793281653745,
757
- "precision_sample_sd": 0.010987847993006481,
758
- "recall_mean": 0.9393939393939394,
759
- "recall_sample_sd": 0.06943296507508849,
760
- "f1_mean": 0.6043956043956044,
761
- "f1_sample_sd": 0.022305256187015846
762
- },
763
- {
764
- "arm": "existing_specialized_separate_lineage",
765
- "label": "Flag-Waving",
766
- "support": 17,
767
- "precision_mean": 0.5576388888888889,
768
- "precision_sample_sd": 0.03705367103320583,
769
- "recall_mean": 0.8823529411764706,
770
- "recall_sample_sd": 0.11764705882352944,
771
- "f1_mean": 0.680769868923179,
772
- "f1_sample_sd": 0.04164658769406446
773
- },
774
- {
775
- "arm": "existing_specialized_separate_lineage",
776
- "label": "Loaded_Language",
777
- "support": 45,
778
- "precision_mean": 0.8387650085763294,
779
- "precision_sample_sd": 0.01782556577257849,
780
- "recall_mean": 1.0,
781
- "recall_sample_sd": 0.0,
782
- "f1_mean": 0.9122448979591837,
783
- "f1_sample_sd": 0.01060439269940129
784
- },
785
- {
786
- "arm": "existing_specialized_separate_lineage",
787
- "label": "Name_Calling,Labeling",
788
- "support": 33,
789
- "precision_mean": 0.6957104190221673,
790
- "precision_sample_sd": 0.019201490809269257,
791
- "recall_mean": 0.9696969696969697,
792
- "recall_sample_sd": 0.030303030303030276,
793
- "f1_mean": 0.8101049442821594,
794
- "f1_sample_sd": 0.022092733328802537
795
- },
796
- {
797
- "arm": "existing_specialized_separate_lineage",
798
- "label": "Repetition",
799
- "support": 23,
800
- "precision_mean": 0.4781557067271353,
801
- "precision_sample_sd": 0.009897987322673213,
802
- "recall_mean": 0.9420289855072463,
803
- "recall_sample_sd": 0.06641414050660639,
804
- "f1_mean": 0.6337774426009719,
805
- "f1_sample_sd": 0.016444118333485254
806
- },
807
- {
808
- "arm": "existing_specialized_separate_lineage",
809
- "label": "Slogans",
810
- "support": 14,
811
- "precision_mean": 0.4334215167548501,
812
- "precision_sample_sd": 0.05811581861282867,
813
- "recall_mean": 0.6904761904761905,
814
- "recall_sample_sd": 0.1649572197684645,
815
- "f1_mean": 0.5201316298877274,
816
- "f1_sample_sd": 0.018567913895656114
817
- },
818
- {
819
- "arm": "existing_specialized_separate_lineage",
820
- "label": "Thought-terminating_Cliches",
821
- "support": 7,
822
- "precision_mean": 0.18446115288220552,
823
- "precision_sample_sd": 0.03641251252864963,
824
- "recall_mean": 0.6666666666666666,
825
- "recall_sample_sd": 0.1649572197684645,
826
- "f1_mean": 0.2868626868626869,
827
- "f1_sample_sd": 0.0511620864429468
828
- },
829
- {
830
- "arm": "existing_specialized_separate_lineage",
831
- "label": "Whataboutism,Straw_Men,Red_Herring",
832
- "support": 10,
833
- "precision_mean": 0.2777777777777778,
834
- "precision_sample_sd": 0.0481125224324688,
835
- "recall_mean": 0.4,
836
- "recall_sample_sd": 0.0,
837
- "f1_mean": 0.32634032634032634,
838
- "f1_sample_sd": 0.03229931575886014
839
- }
840
- ],
841
- "compression_ladder": [
842
- {
843
- "name": "24L upstream text teacher",
844
- "layers": 24,
845
- "parameters": 752393024,
846
- "weight_bytes": 1504825632,
847
- "weight_mib": 1435.1135559082031,
848
- "weight_sha256": "30783da4651259c26b8a204821ea426b9526431e0c14da0d1c6fa284e7414e5e",
849
- "kd_stage": "",
850
- "mean_total_loss": "",
851
- "mean_interface_loss": "",
852
- "mean_final_loss": ""
853
- },
854
- {
855
- "name": "8L task-free KD",
856
- "layers": 8,
857
- "parameters": 420318144,
858
- "weight_bytes": 840647880,
859
- "weight_mib": 801.7042922973633,
860
- "weight_sha256": "086fbaf9a4838ecde56cd3541c9c3212b22d0c5ee745cf2c1c5e0bb31e0ec22c",
861
- "kd_stage": "24to8",
862
- "mean_total_loss": 0.17622653172702485,
863
- "mean_interface_loss": 0.21562811763578793,
864
- "mean_final_loss": 0.16626427527853593
865
- },
866
- {
867
- "name": "6L task-free KD",
868
- "layers": 6,
869
- "parameters": 377207424,
870
- "weight_bytes": 754423424,
871
- "weight_mib": 719.4742431640625,
872
- "weight_sha256": "492749c20f77a42f6bc993f0927a82217b94f872b7eacf25612f7520c9bc28ce",
873
- "kd_stage": "8to6",
874
- "mean_total_loss": 0.03382104352249371,
875
- "mean_interface_loss": 0.04486069045515251,
876
- "mean_final_loss": 0.031147060785087888
877
- },
878
- {
879
- "name": "4L task-free KD root",
880
- "layers": 4,
881
- "parameters": 334096704,
882
- "weight_bytes": 668198976,
883
- "weight_mib": 637.2442016601562,
884
- "weight_sha256": "2732c616772fe320cdea228ab4554981418b1b2bf615c4183fb1ac8e6e2168d3",
885
- "kd_stage": "6to4",
886
- "mean_total_loss": 0.047020394468859195,
887
- "mean_interface_loss": 0.07048421185527332,
888
- "mean_final_loss": 0.043119939065377366
889
- }
890
- ]
891
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
finalization_manifest.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
  "schema_version": "standalone4l-base-finalization-v1",
3
- "source": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/base-4l",
4
- "output": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/base-4l-canonical",
5
  "source_hashes": {
6
  "model.safetensors": "694ce83253652ec1287deaab7cc328ed3ff735ca25e1cef682dca82c80892aa0",
7
  "config.json": "7e95a39a6efe3d34d8d79e27bcebc4068cf91f23b50f4e3bc9a828a0cbb38cf3",
 
1
  {
2
  "schema_version": "standalone4l-base-finalization-v1",
3
+ "source": "results/standalone4l_base/20260920-taskblind-grid/N32768/base-4l",
4
+ "output": "results/standalone4l_base/20260920-taskblind-grid/N32768/base-4l-canonical",
5
  "source_hashes": {
6
  "model.safetensors": "694ce83253652ec1287deaab7cc328ed3ff735ca25e1cef682dca82c80892aa0",
7
  "config.json": "7e95a39a6efe3d34d8d79e27bcebc4068cf91f23b50f4e3bc9a828a0cbb38cf3",
provenance/base/finalization_manifest.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
  "schema_version": "standalone4l-base-finalization-v1",
3
- "source": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/base-4l",
4
- "output": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/base-4l-canonical",
5
  "source_hashes": {
6
  "model.safetensors": "694ce83253652ec1287deaab7cc328ed3ff735ca25e1cef682dca82c80892aa0",
7
  "config.json": "7e95a39a6efe3d34d8d79e27bcebc4068cf91f23b50f4e3bc9a828a0cbb38cf3",
 
1
  {
2
  "schema_version": "standalone4l-base-finalization-v1",
3
+ "source": "results/standalone4l_base/20260920-taskblind-grid/N32768/base-4l",
4
+ "output": "results/standalone4l_base/20260920-taskblind-grid/N32768/base-4l-canonical",
5
  "source_hashes": {
6
  "model.safetensors": "694ce83253652ec1287deaab7cc328ed3ff735ca25e1cef682dca82c80892aa0",
7
  "config.json": "7e95a39a6efe3d34d8d79e27bcebc4068cf91f23b50f4e3bc9a828a0cbb38cf3",
provenance/base/training_manifest.json CHANGED
@@ -3,11 +3,11 @@
3
  "status": "completed",
4
  "timestamp_kst": "2026-09-20T07:56:37.853994+09:00",
5
  "stage": "6to4",
6
- "teacher": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/base-6l",
7
- "student_initial": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/init-4l",
8
- "output": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/base-4l",
9
  "data": {
10
- "path": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/data/public_corpora/wikitext-103-raw-v1-b08601e04326/train-00000-of-00002.parquet",
11
  "sha256": "74da360f23826045b3e6ac6375411fdb15f003030aa74f2596ed08b857cb9212",
12
  "rows_used": 4096,
13
  "labels_read": false,
 
3
  "status": "completed",
4
  "timestamp_kst": "2026-09-20T07:56:37.853994+09:00",
5
  "stage": "6to4",
6
+ "teacher": "results/standalone4l_base/20260920-taskblind-grid/N32768/base-6l",
7
+ "student_initial": "results/standalone4l_base/20260920-taskblind-grid/N32768/init-4l",
8
+ "output": "results/standalone4l_base/20260920-taskblind-grid/N32768/base-4l",
9
  "data": {
10
+ "path": "data/public_corpora/wikitext-103-raw-v1-b08601e04326/train-00000-of-00002.parquet",
11
  "sha256": "74da360f23826045b3e6ac6375411fdb15f003030aa74f2596ed08b857cb9212",
12
  "rows_used": 4096,
13
  "labels_read": false,
provenance/task_agnostic_stages/24to8.json CHANGED
@@ -3,11 +3,11 @@
3
  "status": "completed",
4
  "timestamp_kst": "2026-09-20T07:45:30.667739+09:00",
5
  "stage": "24to8",
6
- "teacher": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/teacher-24l-N32768",
7
- "student_initial": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/init-8l",
8
- "output": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/base-8l",
9
  "data": {
10
- "path": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/data/public_corpora/wikitext-103-raw-v1-b08601e04326/train-00000-of-00002.parquet",
11
  "sha256": "74da360f23826045b3e6ac6375411fdb15f003030aa74f2596ed08b857cb9212",
12
  "rows_used": 4096,
13
  "labels_read": false,
 
3
  "status": "completed",
4
  "timestamp_kst": "2026-09-20T07:45:30.667739+09:00",
5
  "stage": "24to8",
6
+ "teacher": "results/standalone4l_base/20260920-taskblind-grid/teacher-24l-N32768",
7
+ "student_initial": "results/standalone4l_base/20260920-taskblind-grid/N32768/init-8l",
8
+ "output": "results/standalone4l_base/20260920-taskblind-grid/N32768/base-8l",
9
  "data": {
10
+ "path": "data/public_corpora/wikitext-103-raw-v1-b08601e04326/train-00000-of-00002.parquet",
11
  "sha256": "74da360f23826045b3e6ac6375411fdb15f003030aa74f2596ed08b857cb9212",
12
  "rows_used": 4096,
13
  "labels_read": false,
provenance/task_agnostic_stages/6to4.json CHANGED
@@ -3,11 +3,11 @@
3
  "status": "completed",
4
  "timestamp_kst": "2026-09-20T07:56:37.853994+09:00",
5
  "stage": "6to4",
6
- "teacher": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/base-6l",
7
- "student_initial": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/init-4l",
8
- "output": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/base-4l",
9
  "data": {
10
- "path": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/data/public_corpora/wikitext-103-raw-v1-b08601e04326/train-00000-of-00002.parquet",
11
  "sha256": "74da360f23826045b3e6ac6375411fdb15f003030aa74f2596ed08b857cb9212",
12
  "rows_used": 4096,
13
  "labels_read": false,
 
3
  "status": "completed",
4
  "timestamp_kst": "2026-09-20T07:56:37.853994+09:00",
5
  "stage": "6to4",
6
+ "teacher": "results/standalone4l_base/20260920-taskblind-grid/N32768/base-6l",
7
+ "student_initial": "results/standalone4l_base/20260920-taskblind-grid/N32768/init-4l",
8
+ "output": "results/standalone4l_base/20260920-taskblind-grid/N32768/base-4l",
9
  "data": {
10
+ "path": "data/public_corpora/wikitext-103-raw-v1-b08601e04326/train-00000-of-00002.parquet",
11
  "sha256": "74da360f23826045b3e6ac6375411fdb15f003030aa74f2596ed08b857cb9212",
12
  "rows_used": 4096,
13
  "labels_read": false,
provenance/task_agnostic_stages/8to6.json CHANGED
@@ -3,11 +3,11 @@
3
  "status": "completed",
4
  "timestamp_kst": "2026-09-20T07:52:32.717463+09:00",
5
  "stage": "8to6",
6
- "teacher": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/base-8l",
7
- "student_initial": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/init-6l",
8
- "output": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/base-6l",
9
  "data": {
10
- "path": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/data/public_corpora/wikitext-103-raw-v1-b08601e04326/train-00000-of-00002.parquet",
11
  "sha256": "74da360f23826045b3e6ac6375411fdb15f003030aa74f2596ed08b857cb9212",
12
  "rows_used": 4096,
13
  "labels_read": false,
 
3
  "status": "completed",
4
  "timestamp_kst": "2026-09-20T07:52:32.717463+09:00",
5
  "stage": "8to6",
6
+ "teacher": "results/standalone4l_base/20260920-taskblind-grid/N32768/base-8l",
7
+ "student_initial": "results/standalone4l_base/20260920-taskblind-grid/N32768/init-6l",
8
+ "output": "results/standalone4l_base/20260920-taskblind-grid/N32768/base-6l",
9
  "data": {
10
+ "path": "data/public_corpora/wikitext-103-raw-v1-b08601e04326/train-00000-of-00002.parquet",
11
  "sha256": "74da360f23826045b3e6ac6375411fdb15f003030aa74f2596ed08b857cb9212",
12
  "rows_used": 4096,
13
  "labels_read": false,
release_manifest.json DELETED
@@ -1,675 +0,0 @@
1
- {
2
- "schema": "qwen35-unified-classification-base-release-v3-v128k-update",
3
- "status": "complete",
4
- "repo_id": "mp-juuuns/qwen35-standalone4l-classification-base",
5
- "old_repo_preserved": {
6
- "repo_id": "mp-juuuns/qwen35-standalone4l-propaganda-classifier",
7
- "revision": "e7e0ff16828052687d2e8dd7849e7a521629cf38"
8
- },
9
- "root_role": "task_agnostic_headless_classification_backbone",
10
- "root_weight_sha256": "7aa658a1adef9665fa86c0cdd97d0e32c57e3dd6be618eaaf7794448930d387c",
11
- "root_multilingual_metadata_scope": "inherited tokenizer/architecture coverage; no direct multilingual downstream evaluation",
12
- "benchmark_schema": "standalone4l-complete-transfer-benchmark-v2",
13
- "figure_schema": "standalone4l-matplotlib-figures-v1",
14
- "nested_specialized_is_separate_lineage": true,
15
- "source_data_redistributed": false,
16
- "files_before_manifest": [
17
- {
18
- "path": ".gitattributes",
19
- "role": "Hugging Face LFS policy",
20
- "bytes": 50,
21
- "sha256": "ee05350c7fda0af4da116020dd7b976d731de3beed4f5e717d5bcc95ead24550",
22
- "source": "generated"
23
- },
24
- {
25
- "path": "LICENSE",
26
- "role": "Apache-2.0 license",
27
- "bytes": 11418,
28
- "sha256": "77fd4710def9ec3c0f6225800e0235f15a425abd4a8b03559127fcd782612049",
29
- "source": "references/software/transformers-5.13.0/LICENSE"
30
- },
31
- {
32
- "path": "README.md",
33
- "role": "root model card",
34
- "bytes": 16305,
35
- "sha256": "e8456e330ad315f08a6c37e308b6e3bbe54b31f6d65fc37a6a1625386d99a529",
36
- "source": "docs/huggingface/UNIFIED_BASE_RELEASE_MODEL_CARD.md"
37
- },
38
- {
39
- "path": "benchmark/BENCHMARK_CARD.md",
40
- "role": "generated benchmark report",
41
- "bytes": 3173,
42
- "sha256": "bf4073de38c3905effd2dcce6f38894f325c8a0a79bedfb2779f0545c86b9c9f",
43
- "source": "generated"
44
- },
45
- {
46
- "path": "benchmark/arm_summary.csv",
47
- "role": "three-seed metric summary",
48
- "bytes": 1655,
49
- "sha256": "e01bd7ac59a3212e44b77e5fcfbc7c41ec91d74c998306d2438209d6ef1a3aa7",
50
- "source": "generated"
51
- },
52
- {
53
- "path": "benchmark/compression_ladder.csv",
54
- "role": "24L to 4L compression evidence",
55
- "bytes": 827,
56
- "sha256": "2c544c3907c4f592d724f658850b8c6b9634845cb5b50a66ebd2ea6c5d48ce0b",
57
- "source": "generated"
58
- },
59
- {
60
- "path": "benchmark/figures/01_overall_transfer_quality.csv",
61
- "role": "Matplotlib benchmark figures and source data",
62
- "bytes": 500,
63
- "sha256": "769a516e53417e16b61cf30d95a02f6c812c0b1536242cdbf836da779753efbd",
64
- "source": "results/standalone4l_base/20260731-v1/huggingface_figures_v1/01_overall_transfer_quality.csv"
65
- },
66
- {
67
- "path": "benchmark/figures/01_overall_transfer_quality.pdf",
68
- "role": "Matplotlib benchmark figures and source data",
69
- "bytes": 18071,
70
- "sha256": "7221e0d2feb356e4d7f44aa0381e576254f734837e3f54fb0dfb98795ec19ded",
71
- "source": "results/standalone4l_base/20260731-v1/huggingface_figures_v1/01_overall_transfer_quality.pdf"
72
- },
73
- {
74
- "path": "benchmark/figures/01_overall_transfer_quality.png",
75
- "role": "Matplotlib benchmark figures and source data",
76
- "bytes": 82127,
77
- "sha256": "6d35b5964791cfff2701e43831996a9aad7d607305ab3031b0782bed2ffffa66",
78
- "source": "results/standalone4l_base/20260731-v1/huggingface_figures_v1/01_overall_transfer_quality.png"
79
- },
80
- {
81
- "path": "benchmark/figures/02_per_label_transfer_metrics.csv",
82
- "role": "Matplotlib benchmark figures and source data",
83
- "bytes": 1169,
84
- "sha256": "50a3ceb0e3146ba8cade9c86c4b59c7db6cbedfc41fac39a811d7f39f4bdeb0d",
85
- "source": "results/standalone4l_base/20260731-v1/huggingface_figures_v1/02_per_label_transfer_metrics.csv"
86
- },
87
- {
88
- "path": "benchmark/figures/02_per_label_transfer_metrics.pdf",
89
- "role": "Matplotlib benchmark figures and source data",
90
- "bytes": 21379,
91
- "sha256": "10756816e8ad800f5d872907986680abf637e76fe47f9790f27645f8f231ae4a",
92
- "source": "results/standalone4l_base/20260731-v1/huggingface_figures_v1/02_per_label_transfer_metrics.pdf"
93
- },
94
- {
95
- "path": "benchmark/figures/02_per_label_transfer_metrics.png",
96
- "role": "Matplotlib benchmark figures and source data",
97
- "bytes": 138298,
98
- "sha256": "5b558738b405a7b04bb30f38b6e53af21b2392ea12cad0d777e9d6d6985d3d04",
99
- "source": "results/standalone4l_base/20260731-v1/huggingface_figures_v1/02_per_label_transfer_metrics.png"
100
- },
101
- {
102
- "path": "benchmark/figures/03_quality_efficiency.csv",
103
- "role": "Matplotlib benchmark figures and source data",
104
- "bytes": 448,
105
- "sha256": "f2a34cbcd700b2663cf761a5fb1dbd3aa99f310a96d4b4477a3a4529bee1a1b3",
106
- "source": "results/standalone4l_base/20260731-v1/huggingface_figures_v1/03_quality_efficiency.csv"
107
- },
108
- {
109
- "path": "benchmark/figures/03_quality_efficiency.pdf",
110
- "role": "Matplotlib benchmark figures and source data",
111
- "bytes": 21933,
112
- "sha256": "18c81ce50fc7871b000c9e387f0f7d7780002cb2bd249153aadfa55a186342c0",
113
- "source": "results/standalone4l_base/20260731-v1/huggingface_figures_v1/03_quality_efficiency.pdf"
114
- },
115
- {
116
- "path": "benchmark/figures/03_quality_efficiency.png",
117
- "role": "Matplotlib benchmark figures and source data",
118
- "bytes": 99374,
119
- "sha256": "08ed7ca8193d5f9065645fb22491ab88ee260d43566e9013aa0f98d7ab4dde87",
120
- "source": "results/standalone4l_base/20260731-v1/huggingface_figures_v1/03_quality_efficiency.png"
121
- },
122
- {
123
- "path": "benchmark/figures/04_compression_ladder.csv",
124
- "role": "Matplotlib benchmark figures and source data",
125
- "bytes": 352,
126
- "sha256": "2dfac2b8330724ce46a22d8f3ddfeffc767f52b93a93cb1a833c89c5820ffccc",
127
- "source": "results/standalone4l_base/20260731-v1/huggingface_figures_v1/04_compression_ladder.csv"
128
- },
129
- {
130
- "path": "benchmark/figures/04_compression_ladder.pdf",
131
- "role": "Matplotlib benchmark figures and source data",
132
- "bytes": 19490,
133
- "sha256": "953d6e356498127b2a607530fac2e41b07b4bf38ed0af945f778b784dc6f7b23",
134
- "source": "results/standalone4l_base/20260731-v1/huggingface_figures_v1/04_compression_ladder.pdf"
135
- },
136
- {
137
- "path": "benchmark/figures/04_compression_ladder.png",
138
- "role": "Matplotlib benchmark figures and source data",
139
- "bytes": 120809,
140
- "sha256": "9f6c1d6fdefd45711fcfa57a7d0d91f6baaa3a5ee0b44a36554a9c74febe7090",
141
- "source": "results/standalone4l_base/20260731-v1/huggingface_figures_v1/04_compression_ladder.png"
142
- },
143
- {
144
- "path": "benchmark/figures/README.md",
145
- "role": "Matplotlib benchmark figures and source data",
146
- "bytes": 683,
147
- "sha256": "7e00107e6576da117689dca0c164c497ce32e9d93c034331e2e161625afc2092",
148
- "source": "results/standalone4l_base/20260731-v1/huggingface_figures_v1/README.md"
149
- },
150
- {
151
- "path": "benchmark/figures/manifest.json",
152
- "role": "Matplotlib benchmark figures and source data",
153
- "bytes": 640,
154
- "sha256": "41fbd7c746eb52cbb3ae7e7e5427ac1e509de4be65dd441548754e090f6bc74f",
155
- "source": "results/standalone4l_base/20260731-v1/huggingface_figures_v1/manifest.json"
156
- },
157
- {
158
- "path": "benchmark/figures/source_summary.json",
159
- "role": "Matplotlib benchmark figures and source data",
160
- "bytes": 26913,
161
- "sha256": "3d2c9fcf9c6cc7da1a885968b70a42eaa38659d317c9d3290a2b3066f3f9ee6a",
162
- "source": "results/standalone4l_base/20260731-v1/huggingface_figures_v1/source_summary.json"
163
- },
164
- {
165
- "path": "benchmark/figures/v128k_compression.csv",
166
- "role": "generated v128k update artifact",
167
- "source": "generated",
168
- "bytes": 198,
169
- "sha256": "078b4f34897e0a1b9bcff66ce21ba0feb28d72c86eaa8cef608f110c48f52f2e"
170
- },
171
- {
172
- "path": "benchmark/figures/v128k_quality_efficiency.csv",
173
- "role": "generated v128k update artifact",
174
- "source": "generated",
175
- "bytes": 370,
176
- "sha256": "014da12a435e3404b31470e2642d180dc111eb608fb8513fd993097019698b7a"
177
- },
178
- {
179
- "path": "benchmark/figures/v128k_quality_storage.png",
180
- "role": "generated v128k update artifact",
181
- "source": "generated",
182
- "bytes": 73388,
183
- "sha256": "644e5b7469c0b4e0b63d5ecaa1c0f2fc8e07af8bb5d9e4d3ae5a88166de374f3"
184
- },
185
- {
186
- "path": "benchmark/figures/v128k_vocabulary_size.png",
187
- "role": "generated v128k update artifact",
188
- "source": "generated",
189
- "bytes": 52750,
190
- "sha256": "66d5da977653e326893da3d34b5babb2e078ca1381753112e74ef76bace80193"
191
- },
192
- {
193
- "path": "benchmark/per_label_metrics.csv",
194
- "role": "per-label metrics by arm and seed",
195
- "bytes": 12963,
196
- "sha256": "ffc21be32b2106c7bf1df3854097a975ac370da01ac1b732eb5c2ba9075ddeda",
197
- "source": "generated"
198
- },
199
- {
200
- "path": "benchmark/per_label_summary.csv",
201
- "role": "per-label three-seed summary",
202
- "bytes": 7009,
203
- "sha256": "c8e3876abf6cfe960c9332d3bc9d1fbde2f3213f3111351acf83b21bbb184049",
204
- "source": "generated"
205
- },
206
- {
207
- "path": "benchmark/reports/existing_specialized_separate_lineage-seed41.json",
208
- "role": "normalized frozen benchmark report",
209
- "bytes": 115793,
210
- "sha256": "c1bd379680abcf50b47b7d23e4aafbc8b5e7f40a0b0e9700bde21a8ce49af803",
211
- "source": "generated"
212
- },
213
- {
214
- "path": "benchmark/reports/existing_specialized_separate_lineage-seed42.json",
215
- "role": "normalized frozen benchmark report",
216
- "bytes": 115780,
217
- "sha256": "a30e38433d008db4aa7e5fbfd477ab80a0daf50e970f8e7b04397efc1a47843a",
218
- "source": "generated"
219
- },
220
- {
221
- "path": "benchmark/reports/existing_specialized_separate_lineage-seed43.json",
222
- "role": "normalized frozen benchmark report",
223
- "bytes": 115985,
224
- "sha256": "96466290d09d27b8bc044a9f9a95765ce12b4d7b02dfb0123e436e7535623353",
225
- "source": "generated"
226
- },
227
- {
228
- "path": "benchmark/reports/structural_copy_control-seed41.json",
229
- "role": "normalized frozen benchmark report",
230
- "bytes": 115612,
231
- "sha256": "816c3249d3ec281ef73d34dc3ab7576a48415d250c32977eb94ca37f73b0b810",
232
- "source": "generated"
233
- },
234
- {
235
- "path": "benchmark/reports/structural_copy_control-seed42.json",
236
- "role": "normalized frozen benchmark report",
237
- "bytes": 115225,
238
- "sha256": "0b19eabe1e31d0012c754db3b31173f8a6fcdee30973f1a1bebd3a370d11064c",
239
- "source": "generated"
240
- },
241
- {
242
- "path": "benchmark/reports/structural_copy_control-seed43.json",
243
- "role": "normalized frozen benchmark report",
244
- "bytes": 115744,
245
- "sha256": "6dde34173462eac50141ff082c18194f3233f8c06853532d9b12756562e69e2e",
246
- "source": "generated"
247
- },
248
- {
249
- "path": "benchmark/reports/task_agnostic_base-seed41.json",
250
- "role": "normalized frozen benchmark report",
251
- "bytes": 115885,
252
- "sha256": "6b3a2e0372a83e1e71c2eca9596efc8cd1a3e39c28e25e3373f195431b5858a4",
253
- "source": "generated"
254
- },
255
- {
256
- "path": "benchmark/reports/task_agnostic_base-seed42.json",
257
- "role": "normalized frozen benchmark report",
258
- "bytes": 115674,
259
- "sha256": "42dc788983716be0e9847aad87ff79c7e7297884f2635a8d098639be77326639",
260
- "source": "generated"
261
- },
262
- {
263
- "path": "benchmark/reports/task_agnostic_base-seed43.json",
264
- "role": "normalized frozen benchmark report",
265
- "bytes": 115062,
266
- "sha256": "7ee52dc94a152a2f767dcff171f3c9cc1072fabd3ebc7b14db53617b7b1d33d5",
267
- "source": "generated"
268
- },
269
- {
270
- "path": "benchmark/resource_metrics.csv",
271
- "role": "resource measurements by arm and seed",
272
- "bytes": 1649,
273
- "sha256": "bf5cefdec3b2f3646114bd18bbf8e5e39721bde768ba968d73f22e96fc9ca8cf",
274
- "source": "generated"
275
- },
276
- {
277
- "path": "benchmark/seed_metrics.csv",
278
- "role": "overall metrics by arm and seed",
279
- "bytes": 1175,
280
- "sha256": "e0b08a958d25243158a03dff76ca153d2ded76cc3f1771e4046a0068688ad6a8",
281
- "source": "generated"
282
- },
283
- {
284
- "path": "benchmark/semeval-transfer-summary.json",
285
- "role": "compatibility alias for complete transfer summary",
286
- "bytes": 26913,
287
- "sha256": "3d2c9fcf9c6cc7da1a885968b70a42eaa38659d317c9d3290a2b3066f3f9ee6a",
288
- "source": "generated"
289
- },
290
- {
291
- "path": "benchmark/summary.json",
292
- "role": "complete machine-readable transfer benchmark",
293
- "bytes": 26913,
294
- "sha256": "3d2c9fcf9c6cc7da1a885968b70a42eaa38659d317c9d3290a2b3066f3f9ee6a",
295
- "source": "generated"
296
- },
297
- {
298
- "path": "config.json",
299
- "role": "task-agnostic root checkpoint",
300
- "bytes": 2156,
301
- "sha256": "32967cd99c1386f0aecf0822e0e3f67164a72058e709d989b888928b0bbe46c5",
302
- "source": "results/standalone4l_base/20260731-v1/base-4l-canonical/config.json"
303
- },
304
- {
305
- "path": "distillation/README.md",
306
- "role": "reusable public platform",
307
- "bytes": 3432,
308
- "sha256": "f5b7248f7781412e8b5d8f2fb3b0c309568e4e75baf55029b26a6fa985d9d608",
309
- "source": "distillation/README.md"
310
- },
311
- {
312
- "path": "distillation/configs/task_agnostic_base.yaml",
313
- "role": "reusable public platform",
314
- "bytes": 357,
315
- "sha256": "adf8b6e3f0f5a5dfe042dcc09a7a8e7ef8cd1146462b089c5f6c9393e9a3131f",
316
- "source": "distillation/configs/task_agnostic_base.yaml"
317
- },
318
- {
319
- "path": "distillation/examples/labels.json",
320
- "role": "reusable public platform",
321
- "bytes": 36,
322
- "sha256": "8b528ab38eb2e9f53f9ec2c50d4a0c860d87594067dc9b064b4bb47ef51f8f50",
323
- "source": "distillation/examples/labels.json"
324
- },
325
- {
326
- "path": "distillation/examples/multilabel_train.jsonl",
327
- "role": "reusable public platform",
328
- "bytes": 186,
329
- "sha256": "daf5859bb93c474d766ecb0681a2585a6a8ed3cb5efcdd476c845727a0c3583c",
330
- "source": "distillation/examples/multilabel_train.jsonl"
331
- },
332
- {
333
- "path": "distillation/examples/single_label_train.jsonl",
334
- "role": "reusable public platform",
335
- "bytes": 171,
336
- "sha256": "918490e8db364c6e601cfdcb09e28c272eafc8359ffe190645a4cbecdf06aed3",
337
- "source": "distillation/examples/single_label_train.jsonl"
338
- },
339
- {
340
- "path": "distillation/examples/unlabeled.jsonl",
341
- "role": "reusable public platform",
342
- "bytes": 183,
343
- "sha256": "7a1db673674a003ef749b1fa90a99eff6db70c3189411b9cee574cc23d084860",
344
- "source": "distillation/examples/unlabeled.jsonl"
345
- },
346
- {
347
- "path": "distillation/pyproject.toml",
348
- "role": "reusable public platform",
349
- "bytes": 487,
350
- "sha256": "3c5a0f9b48849381307a6b3c258729d4079bf5d42bb6ab5d28edbdda4a1a3592",
351
- "source": "distillation/pyproject.toml"
352
- },
353
- {
354
- "path": "distillation/qwen35_distill/__init__.py",
355
- "role": "reusable public platform",
356
- "bytes": 175,
357
- "sha256": "2965e662310bf4197d1cbd4f9b3fcf3f383950dbfc4f033729a286d026709f3a",
358
- "source": "distillation/qwen35_distill/__init__.py"
359
- },
360
- {
361
- "path": "distillation/qwen35_distill/checkpoint.py",
362
- "role": "reusable public platform",
363
- "bytes": 12541,
364
- "sha256": "d67bf7ed3bfb6cb622d6d35a99a4292f1191395270264e81a7d19283d206e5dc",
365
- "source": "distillation/qwen35_distill/checkpoint.py"
366
- },
367
- {
368
- "path": "distillation/qwen35_distill/classification.py",
369
- "role": "reusable public platform",
370
- "bytes": 14602,
371
- "sha256": "ea06bec682622addc426def4f7e315b3d12f7f595af44f714cc811588cf64782",
372
- "source": "distillation/qwen35_distill/classification.py"
373
- },
374
- {
375
- "path": "distillation/qwen35_distill/cli.py",
376
- "role": "reusable public platform",
377
- "bytes": 8015,
378
- "sha256": "d8fb1584484596f9c35db9c2c296dab08b2dd9bb1b7910a115710feee7562771",
379
- "source": "distillation/qwen35_distill/cli.py"
380
- },
381
- {
382
- "path": "distillation/qwen35_distill/layer_maps.py",
383
- "role": "reusable public platform",
384
- "bytes": 3103,
385
- "sha256": "a9c49af0ee14b307a08c096bcd1ae55fa9bc67bdb807fa9383fc313246e7edd7",
386
- "source": "distillation/qwen35_distill/layer_maps.py"
387
- },
388
- {
389
- "path": "distillation/qwen35_distill/losses.py",
390
- "role": "reusable public platform",
391
- "bytes": 3794,
392
- "sha256": "d9b135cf5b0adf4bbfc321200395fbd6112e49fd1a027b30262bcff1c4c0b654",
393
- "source": "distillation/qwen35_distill/losses.py"
394
- },
395
- {
396
- "path": "distillation/qwen35_distill/schema.py",
397
- "role": "reusable public platform",
398
- "bytes": 5607,
399
- "sha256": "00731ae7838317105f9fff960e3da210b2a55954feb37b3ac793bcfc81fdab2d",
400
- "source": "distillation/qwen35_distill/schema.py"
401
- },
402
- {
403
- "path": "distillation/qwen35_distill/training.py",
404
- "role": "reusable public platform",
405
- "bytes": 10474,
406
- "sha256": "4dc432acb70e03848d2e5964e8550065490fda3857aac32fe3b79895ef72026f",
407
- "source": "distillation/qwen35_distill/training.py"
408
- },
409
- {
410
- "path": "docs/RELEASE_CONTRACT.md",
411
- "role": "release contract",
412
- "bytes": 5381,
413
- "sha256": "1b6e475b9d7d9af52bb8845de020303b1c6e89646d8e4f9be17dbe4ab2910417",
414
- "source": "docs/huggingface/UNIFIED_BASE_RELEASE_CONTRACT.md"
415
- },
416
- {
417
- "path": "model.safetensors",
418
- "role": "task-agnostic root checkpoint",
419
- "bytes": 421783648,
420
- "sha256": "7aa658a1adef9665fa86c0cdd97d0e32c57e3dd6be618eaaf7794448930d387c",
421
- "source": "results/standalone4l_base/20260731-v1/base-4l-canonical/model.safetensors"
422
- },
423
- {
424
- "path": "models/semeval-propaganda/LICENSE",
425
- "role": "byte-identical existing specialized release",
426
- "bytes": 11418,
427
- "sha256": "77fd4710def9ec3c0f6225800e0235f15a425abd4a8b03559127fcd782612049",
428
- "source": "output/huggingface_qwen35_standalone4l_release_v4/LICENSE"
429
- },
430
- {
431
- "path": "models/semeval-propaganda/README.md",
432
- "role": "specialized lineage card",
433
- "bytes": 1403,
434
- "sha256": "bf4ba7cc2514502a28f424507c41c1cfce5089c051ba7500cf46c1d13860722f",
435
- "source": "generated"
436
- },
437
- {
438
- "path": "models/semeval-propaganda/benchmark/final4l_aggregate.json",
439
- "role": "existing specialized aggregate",
440
- "bytes": 10882,
441
- "sha256": "0bae347bb54f8089a5fa5450f8366b5c56eb3e598beaf9a87f4385363a402bcb",
442
- "source": "output/huggingface_qwen35_standalone4l_release_v4/benchmark/final4l_aggregate.json"
443
- },
444
- {
445
- "path": "models/semeval-propaganda/benchmark/seed41.json",
446
- "role": "existing specialized seed report",
447
- "bytes": 115793,
448
- "sha256": "c1bd379680abcf50b47b7d23e4aafbc8b5e7f40a0b0e9700bde21a8ce49af803",
449
- "source": "output/huggingface_qwen35_standalone4l_release_v4/benchmark/seed_reports/seed41.json"
450
- },
451
- {
452
- "path": "models/semeval-propaganda/benchmark/seed42.json",
453
- "role": "existing specialized seed report",
454
- "bytes": 115780,
455
- "sha256": "a30e38433d008db4aa7e5fbfd477ab80a0daf50e970f8e7b04397efc1a47843a",
456
- "source": "output/huggingface_qwen35_standalone4l_release_v4/benchmark/seed_reports/seed42.json"
457
- },
458
- {
459
- "path": "models/semeval-propaganda/benchmark/seed43.json",
460
- "role": "existing specialized seed report",
461
- "bytes": 115985,
462
- "sha256": "96466290d09d27b8bc044a9f9a95765ce12b4d7b02dfb0123e436e7535623353",
463
- "source": "output/huggingface_qwen35_standalone4l_release_v4/benchmark/seed_reports/seed43.json"
464
- },
465
- {
466
- "path": "models/semeval-propaganda/chat_template.jinja",
467
- "role": "byte-identical existing specialized release",
468
- "bytes": 6669,
469
- "sha256": "04b007131663760bf3e581e5a953be77044014e87efe1d2a6ca4b72ec0eac978",
470
- "source": "output/huggingface_qwen35_standalone4l_release_v4/chat_template.jinja"
471
- },
472
- {
473
- "path": "models/semeval-propaganda/classify.py",
474
- "role": "byte-identical existing specialized release",
475
- "bytes": 8070,
476
- "sha256": "1c09956cc5d3553eed0b5f64c20c81064df4373ccfaf4fec26929b8f4819773a",
477
- "source": "output/huggingface_qwen35_standalone4l_release_v4/classify.py"
478
- },
479
- {
480
- "path": "models/semeval-propaganda/config.json",
481
- "role": "byte-identical existing specialized release",
482
- "bytes": 3380,
483
- "sha256": "9f04dfbd7e1696cd3df2ccbe194f5f155efe5baefd1a0b615851f51bf46691a0",
484
- "source": "output/huggingface_qwen35_standalone4l_release_v4/config.json"
485
- },
486
- {
487
- "path": "models/semeval-propaganda/labels.json",
488
- "role": "byte-identical existing specialized release",
489
- "bytes": 891,
490
- "sha256": "946fed4fe4c3d320fd113ec8e5776b10f2707b936f857341bb5443d7a9038248",
491
- "source": "output/huggingface_qwen35_standalone4l_release_v4/labels.json"
492
- },
493
- {
494
- "path": "models/semeval-propaganda/model.safetensors",
495
- "role": "byte-identical existing specialized release",
496
- "bytes": 421812688,
497
- "sha256": "0352da6e4717a2e6f3a9d643d034be2cb1c39632d1bce0ced339ddb62acedb37",
498
- "source": "output/huggingface_qwen35_standalone4l_release_v4/model.safetensors"
499
- },
500
- {
501
- "path": "models/semeval-propaganda/release_manifest.json",
502
- "role": "byte-identical existing specialized release",
503
- "bytes": 30946,
504
- "sha256": "ef64dc628fe9bb4efd2a0b6a739f0c842c6fd4586b1803707e7a071e439cdf69",
505
- "source": "output/huggingface_qwen35_standalone4l_release_v4/release_manifest.json"
506
- },
507
- {
508
- "path": "models/semeval-propaganda/requirements.txt",
509
- "role": "byte-identical existing specialized release",
510
- "bytes": 54,
511
- "sha256": "2994886af208a0120c051745f9c14036e6f2caae8c4af863e5965d5a5bd68fdf",
512
- "source": "output/huggingface_qwen35_standalone4l_release_v4/requirements.txt"
513
- },
514
- {
515
- "path": "models/semeval-propaganda/seeds/seed42/chat_template.jinja",
516
- "role": "byte-identical existing specialized seed",
517
- "bytes": 6669,
518
- "sha256": "04b007131663760bf3e581e5a953be77044014e87efe1d2a6ca4b72ec0eac978",
519
- "source": "output/huggingface_qwen35_standalone4l_release_v4/seeds/seed42/chat_template.jinja"
520
- },
521
- {
522
- "path": "models/semeval-propaganda/seeds/seed42/config.json",
523
- "role": "byte-identical existing specialized seed",
524
- "bytes": 3380,
525
- "sha256": "9f04dfbd7e1696cd3df2ccbe194f5f155efe5baefd1a0b615851f51bf46691a0",
526
- "source": "output/huggingface_qwen35_standalone4l_release_v4/seeds/seed42/config.json"
527
- },
528
- {
529
- "path": "models/semeval-propaganda/seeds/seed42/model.safetensors",
530
- "role": "byte-identical existing specialized seed",
531
- "bytes": 421812688,
532
- "sha256": "2db5f4be391d9dee64458cb5865530f39747019e7a219443b5f7850befbce697",
533
- "source": "output/huggingface_qwen35_standalone4l_release_v4/seeds/seed42/model.safetensors"
534
- },
535
- {
536
- "path": "models/semeval-propaganda/seeds/seed42/tokenizer.json",
537
- "role": "byte-identical existing specialized seed",
538
- "bytes": 9223900,
539
- "sha256": "dd2b27bbd4f5f650c7200300013ef03ee01efe620b128343163521bf04806fc8",
540
- "source": "output/huggingface_qwen35_standalone4l_release_v4/seeds/seed42/tokenizer.json"
541
- },
542
- {
543
- "path": "models/semeval-propaganda/seeds/seed42/tokenizer_config.json",
544
- "role": "byte-identical existing specialized seed",
545
- "bytes": 1314,
546
- "sha256": "2215318fe27878ec67e772931bb5f40a98658348d33f2672a575f766a2190d35",
547
- "source": "output/huggingface_qwen35_standalone4l_release_v4/seeds/seed42/tokenizer_config.json"
548
- },
549
- {
550
- "path": "models/semeval-propaganda/seeds/seed43/chat_template.jinja",
551
- "role": "byte-identical existing specialized seed",
552
- "bytes": 6669,
553
- "sha256": "04b007131663760bf3e581e5a953be77044014e87efe1d2a6ca4b72ec0eac978",
554
- "source": "output/huggingface_qwen35_standalone4l_release_v4/seeds/seed43/chat_template.jinja"
555
- },
556
- {
557
- "path": "models/semeval-propaganda/seeds/seed43/config.json",
558
- "role": "byte-identical existing specialized seed",
559
- "bytes": 3380,
560
- "sha256": "9f04dfbd7e1696cd3df2ccbe194f5f155efe5baefd1a0b615851f51bf46691a0",
561
- "source": "output/huggingface_qwen35_standalone4l_release_v4/seeds/seed43/config.json"
562
- },
563
- {
564
- "path": "models/semeval-propaganda/seeds/seed43/model.safetensors",
565
- "role": "byte-identical existing specialized seed",
566
- "bytes": 421812688,
567
- "sha256": "81a56d497df758e8a292eea528438cfc4d5a847cd37d3a7e1107fefa7c500640",
568
- "source": "output/huggingface_qwen35_standalone4l_release_v4/seeds/seed43/model.safetensors"
569
- },
570
- {
571
- "path": "models/semeval-propaganda/seeds/seed43/tokenizer.json",
572
- "role": "byte-identical existing specialized seed",
573
- "bytes": 9223900,
574
- "sha256": "dd2b27bbd4f5f650c7200300013ef03ee01efe620b128343163521bf04806fc8",
575
- "source": "output/huggingface_qwen35_standalone4l_release_v4/seeds/seed43/tokenizer.json"
576
- },
577
- {
578
- "path": "models/semeval-propaganda/seeds/seed43/tokenizer_config.json",
579
- "role": "byte-identical existing specialized seed",
580
- "bytes": 1314,
581
- "sha256": "2215318fe27878ec67e772931bb5f40a98658348d33f2672a575f766a2190d35",
582
- "source": "output/huggingface_qwen35_standalone4l_release_v4/seeds/seed43/tokenizer_config.json"
583
- },
584
- {
585
- "path": "models/semeval-propaganda/thresholds.json",
586
- "role": "byte-identical existing specialized release",
587
- "bytes": 2631,
588
- "sha256": "737c4e145a5b51660c4c45a8acea4c32219cbbbc0706abe235e36b2ad430f17d",
589
- "source": "output/huggingface_qwen35_standalone4l_release_v4/thresholds.json"
590
- },
591
- {
592
- "path": "models/semeval-propaganda/tokenizer.json",
593
- "role": "byte-identical existing specialized release",
594
- "bytes": 9223900,
595
- "sha256": "dd2b27bbd4f5f650c7200300013ef03ee01efe620b128343163521bf04806fc8",
596
- "source": "output/huggingface_qwen35_standalone4l_release_v4/tokenizer.json"
597
- },
598
- {
599
- "path": "models/semeval-propaganda/tokenizer_config.json",
600
- "role": "byte-identical existing specialized release",
601
- "bytes": 1314,
602
- "sha256": "2215318fe27878ec67e772931bb5f40a98658348d33f2672a575f766a2190d35",
603
- "source": "output/huggingface_qwen35_standalone4l_release_v4/tokenizer_config.json"
604
- },
605
- {
606
- "path": "provenance/base/finalization_manifest.json",
607
- "role": "normalized base provenance",
608
- "bytes": 1830,
609
- "sha256": "e831b8a121fe896bb9f033287a5b82a0bffad1037a45bbb1ec913fed261add92",
610
- "source": "generated"
611
- },
612
- {
613
- "path": "provenance/base/training_manifest.json",
614
- "role": "normalized base provenance",
615
- "bytes": 3612,
616
- "sha256": "ec734855870758cbfef99031dac2ded85b54c496532e5c514a9a4f63287c3493",
617
- "source": "generated"
618
- },
619
- {
620
- "path": "provenance/task_agnostic_stages/24to8.json",
621
- "role": "normalized task-agnostic KD stage manifest",
622
- "bytes": 3596,
623
- "sha256": "b9ce99b993c4f4563bc2521fc731bfd82f5e8149c141133599fbd8f982ba43d9",
624
- "source": "generated"
625
- },
626
- {
627
- "path": "provenance/task_agnostic_stages/6to4.json",
628
- "role": "normalized task-agnostic KD stage manifest",
629
- "bytes": 3612,
630
- "sha256": "ec734855870758cbfef99031dac2ded85b54c496532e5c514a9a4f63287c3493",
631
- "source": "generated"
632
- },
633
- {
634
- "path": "provenance/task_agnostic_stages/8to6.json",
635
- "role": "normalized task-agnostic KD stage manifest",
636
- "bytes": 3611,
637
- "sha256": "65e5383afbbcfe0b7dd78a5411c319ed6bf2586659fb1892aca9ecbd5dfd4993",
638
- "source": "generated"
639
- },
640
- {
641
- "path": "requirements.txt",
642
- "role": "root runtime requirements",
643
- "bytes": 49,
644
- "sha256": "7b48e7811e56e552fc4aef44189e1d7c9b1352c85d2b3f10ff0bdd1ea5cdd824",
645
- "source": "generated"
646
- },
647
- {
648
- "path": "tokenizer.json",
649
- "role": "task-agnostic root checkpoint",
650
- "bytes": 9223900,
651
- "sha256": "dd2b27bbd4f5f650c7200300013ef03ee01efe620b128343163521bf04806fc8",
652
- "source": "results/standalone4l_base/20260731-v1/base-4l-canonical/tokenizer.json"
653
- },
654
- {
655
- "path": "tokenizer_config.json",
656
- "role": "task-agnostic root checkpoint",
657
- "bytes": 1314,
658
- "sha256": "2215318fe27878ec67e772931bb5f40a98658348d33f2672a575f766a2190d35",
659
- "source": "results/standalone4l_base/20260731-v1/base-4l-canonical/tokenizer_config.json"
660
- },
661
- {
662
- "path": "v128k_remap_manifest.json",
663
- "role": "generated v128k update artifact",
664
- "source": "generated",
665
- "bytes": 1068,
666
- "sha256": "96afea013b2da25de347afdd87006ae3fdde55328847b9d5186408211d9d8f9a"
667
- }
668
- ],
669
- "root_vocab_size": 128000,
670
- "root_parameter_count": 210889024,
671
- "root_update": {
672
- "type": "token-string-embedding-remap",
673
- "quality_status": "NR for new headless root; existing transfer evidence is descriptive and lineage-confounded"
674
- }
675
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
training_manifest.json CHANGED
@@ -3,11 +3,11 @@
3
  "status": "completed",
4
  "timestamp_kst": "2026-09-20T07:56:37.853994+09:00",
5
  "stage": "6to4",
6
- "teacher": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/base-6l",
7
- "student_initial": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/init-4l",
8
- "output": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/results/standalone4l_base/20260920-taskblind-grid/N32768/base-4l",
9
  "data": {
10
- "path": "/home/mp_ubuntu/Projects/paper_qwen_intent_classifier/data/public_corpora/wikitext-103-raw-v1-b08601e04326/train-00000-of-00002.parquet",
11
  "sha256": "74da360f23826045b3e6ac6375411fdb15f003030aa74f2596ed08b857cb9212",
12
  "rows_used": 4096,
13
  "labels_read": false,
 
3
  "status": "completed",
4
  "timestamp_kst": "2026-09-20T07:56:37.853994+09:00",
5
  "stage": "6to4",
6
+ "teacher": "results/standalone4l_base/20260920-taskblind-grid/N32768/base-6l",
7
+ "student_initial": "results/standalone4l_base/20260920-taskblind-grid/N32768/init-4l",
8
+ "output": "results/standalone4l_base/20260920-taskblind-grid/N32768/base-4l",
9
  "data": {
10
+ "path": "data/public_corpora/wikitext-103-raw-v1-b08601e04326/train-00000-of-00002.parquet",
11
  "sha256": "74da360f23826045b3e6ac6375411fdb15f003030aa74f2596ed08b857cb9212",
12
  "rows_used": 4096,
13
  "labels_read": false,