Add Qwen3.5-2B seed-variation replicates (basis-stability appendix) with config/metrics sidecars; update manifest and README

README.md

@@ -123,6 +123,30 @@ paper's default. Decomposing against a different preprocessing breaks the
 identity. See the GitHub README quickstart for the full decomposition
 snippet.
 
+## Seed-variation replicates (basis-stability appendix)
+
+Six extra Qwen3.5-2B dictionaries, three independently seeded runs per
+width, used end to end by the paper's basis-stability appendix (cross-seed
+stability of contrast explanations, feature-group matching, leave-one-out
+core recovery). They were trained under the seed-variation preprocessing
+recipe, which the paper notes differs from the recipe behind the released
+`qwen3.5-2b/k128_16x` and `k256_32x` dictionaries above, so seed 0 here is
+not the released checkpoint and the appendix never mixes the two families.
+Each directory also carries the run's `config.yaml` and `metrics.json`.
+Same layout and loader as the main checkpoints.
+
+| Path | seed | width | d_features | k | rowEV | top-1 | KL (bits) |
+|---|---:|---|---:|---:|---:|---:|---:|
+| `qwen3.5-2b/k128_16x_seedvar_s0` | 0 | 16x | 32768 | 128 | 0.744 | 0.805 | 0.286 |
+| `qwen3.5-2b/k128_16x_seedvar_s1` | 1 | 16x | 32768 | 128 | 0.745 | 0.805 | 0.291 |
+| `qwen3.5-2b/k128_16x_seedvar_s2` | 2 | 16x | 32768 | 128 | 0.744 | 0.816 | 0.293 |
+| `qwen3.5-2b/k256_32x_seedvar_s0` | 0 | 32x | 65536 | 256 | 0.824 | 0.807 | 0.202 |
+| `qwen3.5-2b/k256_32x_seedvar_s1` | 1 | 32x | 65536 | 256 | 0.823 | 0.822 | 0.222 |
+| `qwen3.5-2b/k256_32x_seedvar_s2` | 2 | 32x | 65536 | 256 | 0.826 | 0.848 | 0.194 |
+
+`rowEV`, `top-1` and `KL` are the held-out training-time eval stored in each
+`metrics.json` (`row_centered_ev`, `val_top1_match`, `val_logit_kl_bits_mean`).
+
 ## Intended use & limitations
 
 Research artifact for mechanistic interpretability of the **final readout**.

manifest.json

@@ -319,5 +319,154 @@
    "layout": "flat",
    "metrics_source": "checkpoint['metrics'] (held-out eval); paper Appendix E"
   }
- ]
-}
+ ],
+ "replicates": {
+  "note": "Seed-variation family used end to end by the paper's basis-stability appendix (cross-seed stability, feature-group matching, leave-one-out core recovery). Three independently seeded dictionaries per width, trained on Qwen3.5-2B under the seed-variation preprocessing recipe. The paper says this recipe differs from the one behind the released qwen3.5-2b/k128_16x and k256_32x dictionaries, so seed 0 here is NOT the released checkpoint; no table in that appendix mixes the two families.",
+  "checkpoints": [
+   {
+    "id": "qwen2b_k128_16x_seedvar_s0",
+    "path": "qwen3.5-2b/k128_16x_seedvar_s0/checkpoint.pt",
+    "model": "Qwen3.5-2B",
+    "base_model": "Qwen/Qwen3.5-2B",
+    "family": "seed_variation",
+    "seed": 0,
+    "width": "16x",
+    "d_features": 32768,
+    "d_model": 2048,
+    "k": 128,
+    "rowEV": 0.744,
+    "top1": 0.805,
+    "KL_bits": 0.286,
+    "steps": 15000,
+    "size_bytes": 539994553,
+    "sha256": "cd5462a450f2696c8353a3b03e2b222ee57f8e0c912473b8850f71a6eada99ed",
+    "layout": "config",
+    "metrics_source": "metrics.json alongside the checkpoint (held-out eval at training time)",
+    "sidecars": [
+     "qwen3.5-2b/k128_16x_seedvar_s0/config.yaml",
+     "qwen3.5-2b/k128_16x_seedvar_s0/metrics.json"
+    ]
+   },
+   {
+    "id": "qwen2b_k128_16x_seedvar_s1",
+    "path": "qwen3.5-2b/k128_16x_seedvar_s1/checkpoint.pt",
+    "model": "Qwen3.5-2B",
+    "base_model": "Qwen/Qwen3.5-2B",
+    "family": "seed_variation",
+    "seed": 1,
+    "width": "16x",
+    "d_features": 32768,
+    "d_model": 2048,
+    "k": 128,
+    "rowEV": 0.745,
+    "top1": 0.805,
+    "KL_bits": 0.291,
+    "steps": 20000,
+    "size_bytes": 539994553,
+    "sha256": "450c99237a3252084f43d12c6eda1f41a4eaf56ac2391ccbd9d8c9c34d95b584",
+    "layout": "config",
+    "metrics_source": "metrics.json alongside the checkpoint (held-out eval at training time)",
+    "sidecars": [
+     "qwen3.5-2b/k128_16x_seedvar_s1/config.yaml",
+     "qwen3.5-2b/k128_16x_seedvar_s1/metrics.json"
+    ]
+   },
+   {
+    "id": "qwen2b_k128_16x_seedvar_s2",
+    "path": "qwen3.5-2b/k128_16x_seedvar_s2/checkpoint.pt",
+    "model": "Qwen3.5-2B",
+    "base_model": "Qwen/Qwen3.5-2B",
+    "family": "seed_variation",
+    "seed": 2,
+    "width": "16x",
+    "d_features": 32768,
+    "d_model": 2048,
+    "k": 128,
+    "rowEV": 0.744,
+    "top1": 0.816,
+    "KL_bits": 0.293,
+    "steps": 20000,
+    "size_bytes": 539994553,
+    "sha256": "485432a34dd5b9a135487d5358cedd319197489dc1d493bcf801ce0e55ec80aa",
+    "layout": "config",
+    "metrics_source": "metrics.json alongside the checkpoint (held-out eval at training time)",
+    "sidecars": [
+     "qwen3.5-2b/k128_16x_seedvar_s2/config.yaml",
+     "qwen3.5-2b/k128_16x_seedvar_s2/metrics.json"
+    ]
+   },
+   {
+    "id": "qwen2b_k256_32x_seedvar_s0",
+    "path": "qwen3.5-2b/k256_32x_seedvar_s0/checkpoint.pt",
+    "model": "Qwen3.5-2B",
+    "base_model": "Qwen/Qwen3.5-2B",
+    "family": "seed_variation",
+    "seed": 0,
+    "width": "32x",
+    "d_features": 65536,
+    "d_model": 2048,
+    "k": 256,
+    "rowEV": 0.824,
+    "top1": 0.807,
+    "KL_bits": 0.202,
+    "steps": 17500,
+    "size_bytes": 1076996537,
+    "sha256": "81cad1148e0ed0d9065691b5fec2dc033f51a5374f8b23d0c1c01be830d653e7",
+    "layout": "config",
+    "metrics_source": "metrics.json alongside the checkpoint (held-out eval at training time)",
+    "sidecars": [
+     "qwen3.5-2b/k256_32x_seedvar_s0/config.yaml",
+     "qwen3.5-2b/k256_32x_seedvar_s0/metrics.json"
+    ]
+   },
+   {
+    "id": "qwen2b_k256_32x_seedvar_s1",
+    "path": "qwen3.5-2b/k256_32x_seedvar_s1/checkpoint.pt",
+    "model": "Qwen3.5-2B",
+    "base_model": "Qwen/Qwen3.5-2B",
+    "family": "seed_variation",
+    "seed": 1,
+    "width": "32x",
+    "d_features": 65536,
+    "d_model": 2048,
+    "k": 256,
+    "rowEV": 0.823,
+    "top1": 0.822,
+    "KL_bits": 0.222,
+    "steps": 17500,
+    "size_bytes": 1076996537,
+    "sha256": "f0ce4816b260a25e23834472ae76b3c4a01b1bcb18a681a21f16349aa6cd9acf",
+    "layout": "config",
+    "metrics_source": "metrics.json alongside the checkpoint (held-out eval at training time)",
+    "sidecars": [
+     "qwen3.5-2b/k256_32x_seedvar_s1/config.yaml",
+     "qwen3.5-2b/k256_32x_seedvar_s1/metrics.json"
+    ]
+   },
+   {
+    "id": "qwen2b_k256_32x_seedvar_s2",
+    "path": "qwen3.5-2b/k256_32x_seedvar_s2/checkpoint.pt",
+    "model": "Qwen3.5-2B",
+    "base_model": "Qwen/Qwen3.5-2B",
+    "family": "seed_variation",
+    "seed": 2,
+    "width": "32x",
+    "d_features": 65536,
+    "d_model": 2048,
+    "k": 256,
+    "rowEV": 0.826,
+    "top1": 0.848,
+    "KL_bits": 0.194,
+    "steps": 17500,
+    "size_bytes": 1076996537,
+    "sha256": "c30f537c387c66f5a92fa5321aef38278d69ede42b9f78254e288160735d7a57",
+    "layout": "config",
+    "metrics_source": "metrics.json alongside the checkpoint (held-out eval at training time)",
+    "sidecars": [
+     "qwen3.5-2b/k256_32x_seedvar_s2/config.yaml",
+     "qwen3.5-2b/k256_32x_seedvar_s2/metrics.json"
+    ]
+   }
+  ]
+ }
+}

qwen3.5-2b/k128_16x_seedvar_s0/checkpoint.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd5462a450f2696c8353a3b03e2b222ee57f8e0c912473b8850f71a6eada99ed
+size 539994553

qwen3.5-2b/k128_16x_seedvar_s0/config.yaml

@@ -0,0 +1,63 @@
+run:
+  name: qwen2b_d32768_k128_s0
+  seed: 0
+  init_seed: 0
+  output_dir: results/exp5_seedvar/qwen2b_d32768_k128_s0
+  device: auto
+data:
+  data_seed: 0
+  fallback: error
+  max_hidden: 2048
+  max_rows: null
+  path: /nfs-share/mh2274/readout_prism_workspace/data/qwen35-2b/qwen35_2b.pt
+  row_preprocessing: center_normalize
+  val_hidden: 512
+factorizer:
+  architecture: topk
+  d_features: 32768
+  encoder_init_scale: 0.02
+  k: 128
+training:
+  auxk_alpha: 0.0
+  auxk_dead_window: 1000000
+  auxk_k: 1024
+  batch_size: 4096
+  code_l2: 1.0e-06
+  final_lr_frac: 0.1
+  init_mode: row_seeded
+  l0_coeff: 0.0
+  l0_controller:
+    coeff_init: 0.0
+    coeff_max: 0.01
+    coeff_min: 1.0e-08
+    control_period: 10
+    ema: 0.9
+    enabled: false
+    err_clip: 0.3
+    kp: 0.2
+    target_l0: 0
+    warmup_steps: 200
+  lambda_mode: delayed_ramp
+  lambda_prism: 0.001
+  lambda_ramp_shape: linear
+  lambda_ramp_start_frac: 0.7
+  lr: 0.001
+  lr_schedule: warmup_cosine
+  prism_top_n: 1
+  row_sampling: hybrid_50freq_50uniform
+  steps: 20000
+  warmup_steps: 1000
+  weight_decay: 0.0001
+evaluation:
+  k: 128
+  max_eval_rows: 20000
+  max_eval_hidden: 1024
+  topk_overlap: 5
+  contribution_top_k: 8
+artifacts:
+  save_checkpoint: true
+  make_figures: false
+autoresearch:
+  score_thresholds:
+    max_dead_feature_rate: 0.5
+    max_val_logit_kl_bits_mean: 20.0

qwen3.5-2b/k128_16x_seedvar_s0/metrics.json

@@ -0,0 +1,39 @@
+{
+  "always_on_feature_rate": 3.0517578125e-05,
+  "architecture": "topk",
+  "d_features": 32768,
+  "data_seed": 0,
+  "data_source": "/nfs-share/mh2274/readout_prism_workspace/data/qwen35-2b/qwen35_2b.pt",
+  "dead_feature_rate": 0.22991943359375,
+  "device": "cuda",
+  "duplicate_decoder_frac": 0.0,
+  "eval_hidden": 512,
+  "eval_rows": 20000,
+  "feature_usage_entropy": 0.8923149172784841,
+  "finite_metrics": true,
+  "identity_abs_error_max": 1.52587890625e-05,
+  "identity_abs_error_mean": 2.9578804969787598e-06,
+  "init_seed": 0,
+  "k": 128,
+  "l0_controller_enabled": false,
+  "l0_controller_final_coeff": null,
+  "l0_controller_target": null,
+  "l0_p10": 128.0,
+  "l0_p50": 128.0,
+  "l0_p90": 128.0,
+  "mean_l0": 127.92695,
+  "mean_l0_train": 127.33492410909403,
+  "rare_feature_rate": 0.436737060546875,
+  "row_centered_cosine": 0.8603649139404297,
+  "row_centered_ev": 0.7441627979278564,
+  "selection_score": 3.0189160955919947,
+  "top1_logit_residual_abs_mean": 0.6984108686447144,
+  "top1_logit_residual_frac_mean": 0.03981868550181389,
+  "top8_abs_contrib_coverage_mean": 0.8226097822189331,
+  "top8_positive_contrib_coverage_mean": 0.8565649390220642,
+  "train_seconds": 5655.247106075287,
+  "used_fallback_data": false,
+  "val_logit_kl_bits_mean": 0.2864432103961134,
+  "val_top1_match": 0.8046875,
+  "val_top5_overlap": 0.7937500000000036
+}

qwen3.5-2b/k128_16x_seedvar_s1/checkpoint.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:450c99237a3252084f43d12c6eda1f41a4eaf56ac2391ccbd9d8c9c34d95b584
+size 539994553

qwen3.5-2b/k128_16x_seedvar_s1/config.yaml

@@ -0,0 +1,63 @@
+run:
+  name: qwen2b_d32768_k128_s1
+  seed: 1
+  init_seed: 1
+  output_dir: results/exp5_seedvar/qwen2b_d32768_k128_s1
+  device: auto
+data:
+  data_seed: 1
+  fallback: error
+  max_hidden: 2048
+  max_rows: null
+  path: /nfs-share/mh2274/readout_prism_workspace/data/qwen35-2b/qwen35_2b.pt
+  row_preprocessing: center_normalize
+  val_hidden: 512
+factorizer:
+  architecture: topk
+  d_features: 32768
+  encoder_init_scale: 0.02
+  k: 128
+training:
+  auxk_alpha: 0.0
+  auxk_dead_window: 1000000
+  auxk_k: 1024
+  batch_size: 4096
+  code_l2: 1.0e-06
+  final_lr_frac: 0.1
+  init_mode: row_seeded
+  l0_coeff: 0.0
+  l0_controller:
+    coeff_init: 0.0
+    coeff_max: 0.01
+    coeff_min: 1.0e-08
+    control_period: 10
+    ema: 0.9
+    enabled: false
+    err_clip: 0.3
+    kp: 0.2
+    target_l0: 0
+    warmup_steps: 200
+  lambda_mode: delayed_ramp
+  lambda_prism: 0.001
+  lambda_ramp_shape: linear
+  lambda_ramp_start_frac: 0.7
+  lr: 0.001
+  lr_schedule: warmup_cosine
+  prism_top_n: 1
+  row_sampling: hybrid_50freq_50uniform
+  steps: 20000
+  warmup_steps: 1000
+  weight_decay: 0.0001
+evaluation:
+  k: 128
+  max_eval_rows: 20000
+  max_eval_hidden: 1024
+  topk_overlap: 5
+  contribution_top_k: 8
+artifacts:
+  save_checkpoint: true
+  make_figures: false
+autoresearch:
+  score_thresholds:
+    max_dead_feature_rate: 0.5
+    max_val_logit_kl_bits_mean: 20.0

qwen3.5-2b/k128_16x_seedvar_s1/metrics.json

@@ -0,0 +1,39 @@
+{
+  "always_on_feature_rate": 3.0517578125e-05,
+  "architecture": "topk",
+  "d_features": 32768,
+  "data_seed": 1,
+  "data_source": "/nfs-share/mh2274/readout_prism_workspace/data/qwen35-2b/qwen35_2b.pt",
+  "dead_feature_rate": 0.226043701171875,
+  "device": "cuda",
+  "duplicate_decoder_frac": 0.0,
+  "eval_hidden": 512,
+  "eval_rows": 20000,
+  "feature_usage_entropy": 0.8926494129565465,
+  "finite_metrics": true,
+  "identity_abs_error_max": 1.52587890625e-05,
+  "identity_abs_error_mean": 3.0249357223510742e-06,
+  "init_seed": 1,
+  "k": 128,
+  "l0_controller_enabled": false,
+  "l0_controller_final_coeff": null,
+  "l0_controller_target": null,
+  "l0_p10": 128.0,
+  "l0_p50": 128.0,
+  "l0_p90": 128.0,
+  "mean_l0": 127.9174,
+  "mean_l0_train": 127.28619622564935,
+  "rare_feature_rate": 0.43365478515625,
+  "row_centered_cosine": 0.8606857657432556,
+  "row_centered_ev": 0.7446612417697906,
+  "selection_score": 3.0473394159174134,
+  "top1_logit_residual_abs_mean": 0.5831794738769531,
+  "top1_logit_residual_frac_mean": 0.0326199010014534,
+  "top8_abs_contrib_coverage_mean": 0.8390382528305054,
+  "top8_positive_contrib_coverage_mean": 0.870077908039093,
+  "train_seconds": 5669.302875995636,
+  "used_fallback_data": false,
+  "val_logit_kl_bits_mean": 0.2912036256785019,
+  "val_top1_match": 0.8046875,
+  "val_top5_overlap": 0.8078125000000034
+}

qwen3.5-2b/k128_16x_seedvar_s2/checkpoint.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:485432a34dd5b9a135487d5358cedd319197489dc1d493bcf801ce0e55ec80aa
+size 539994553

qwen3.5-2b/k128_16x_seedvar_s2/config.yaml

@@ -0,0 +1,63 @@
+run:
+  name: qwen2b_d32768_k128_s2
+  seed: 2
+  init_seed: 2
+  output_dir: results/exp5_seedvar/qwen2b_d32768_k128_s2
+  device: auto
+data:
+  data_seed: 2
+  fallback: error
+  max_hidden: 2048
+  max_rows: null
+  path: /nfs-share/mh2274/readout_prism_workspace/data/qwen35-2b/qwen35_2b.pt
+  row_preprocessing: center_normalize
+  val_hidden: 512
+factorizer:
+  architecture: topk
+  d_features: 32768
+  encoder_init_scale: 0.02
+  k: 128
+training:
+  auxk_alpha: 0.0
+  auxk_dead_window: 1000000
+  auxk_k: 1024
+  batch_size: 4096
+  code_l2: 1.0e-06
+  final_lr_frac: 0.1
+  init_mode: row_seeded
+  l0_coeff: 0.0
+  l0_controller:
+    coeff_init: 0.0
+    coeff_max: 0.01
+    coeff_min: 1.0e-08
+    control_period: 10
+    ema: 0.9
+    enabled: false
+    err_clip: 0.3
+    kp: 0.2
+    target_l0: 0
+    warmup_steps: 200
+  lambda_mode: delayed_ramp
+  lambda_prism: 0.001
+  lambda_ramp_shape: linear
+  lambda_ramp_start_frac: 0.7
+  lr: 0.001
+  lr_schedule: warmup_cosine
+  prism_top_n: 1
+  row_sampling: hybrid_50freq_50uniform
+  steps: 20000
+  warmup_steps: 1000
+  weight_decay: 0.0001
+evaluation:
+  k: 128
+  max_eval_rows: 20000
+  max_eval_hidden: 1024
+  topk_overlap: 5
+  contribution_top_k: 8
+artifacts:
+  save_checkpoint: true
+  make_figures: false
+autoresearch:
+  score_thresholds:
+    max_dead_feature_rate: 0.5
+    max_val_logit_kl_bits_mean: 20.0

qwen3.5-2b/k128_16x_seedvar_s2/metrics.json

@@ -0,0 +1,39 @@
+{
+  "always_on_feature_rate": 3.0517578125e-05,
+  "architecture": "topk",
+  "d_features": 32768,
+  "data_seed": 2,
+  "data_source": "/nfs-share/mh2274/readout_prism_workspace/data/qwen35-2b/qwen35_2b.pt",
+  "dead_feature_rate": 0.223114013671875,
+  "device": "cuda",
+  "duplicate_decoder_frac": 0.0,
+  "eval_hidden": 512,
+  "eval_rows": 20000,
+  "feature_usage_entropy": 0.892924625597383,
+  "finite_metrics": true,
+  "identity_abs_error_max": 1.71661376953125e-05,
+  "identity_abs_error_mean": 2.911314368247986e-06,
+  "init_seed": 2,
+  "k": 128,
+  "l0_controller_enabled": false,
+  "l0_controller_final_coeff": null,
+  "l0_controller_target": null,
+  "l0_p10": 128.0,
+  "l0_p50": 128.0,
+  "l0_p90": 128.0,
+  "mean_l0": 127.92345,
+  "mean_l0_train": 127.30729865837287,
+  "rare_feature_rate": 0.4349365234375,
+  "row_centered_cosine": 0.8605742454528809,
+  "row_centered_ev": 0.7444726824760437,
+  "selection_score": 3.038355168867404,
+  "top1_logit_residual_abs_mean": 0.6107032299041748,
+  "top1_logit_residual_frac_mean": 0.03449808061122894,
+  "top8_abs_contrib_coverage_mean": 0.8264878988265991,
+  "top8_positive_contrib_coverage_mean": 0.8593701720237732,
+  "train_seconds": 5417.30718421936,
+  "used_fallback_data": false,
+  "val_logit_kl_bits_mean": 0.292914444595791,
+  "val_top1_match": 0.81640625,
+  "val_top5_overlap": 0.8062500000000037
+}

qwen3.5-2b/k256_32x_seedvar_s0/checkpoint.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81cad1148e0ed0d9065691b5fec2dc033f51a5374f8b23d0c1c01be830d653e7
+size 1076996537

qwen3.5-2b/k256_32x_seedvar_s0/config.yaml

@@ -0,0 +1,63 @@
+run:
+  name: qwen2b_d65536_k256_s0
+  seed: 0
+  init_seed: 0
+  output_dir: results/exp5_seedvar/qwen2b_d65536_k256_s0
+  device: auto
+data:
+  data_seed: 0
+  fallback: error
+  max_hidden: 2048
+  max_rows: null
+  path: /nfs-share/mh2274/readout_prism_workspace/data/qwen35-2b/qwen35_2b.pt
+  row_preprocessing: center_normalize
+  val_hidden: 512
+factorizer:
+  architecture: topk
+  d_features: 65536
+  encoder_init_scale: 0.02
+  k: 256
+training:
+  auxk_alpha: 0.0
+  auxk_dead_window: 1000000
+  auxk_k: 1024
+  batch_size: 4096
+  code_l2: 1.0e-06
+  final_lr_frac: 0.1
+  init_mode: row_seeded
+  l0_coeff: 0.0
+  l0_controller:
+    coeff_init: 0.0
+    coeff_max: 0.01
+    coeff_min: 1.0e-08
+    control_period: 10
+    ema: 0.9
+    enabled: false
+    err_clip: 0.3
+    kp: 0.2
+    target_l0: 0
+    warmup_steps: 200
+  lambda_mode: delayed_ramp
+  lambda_prism: 0.001
+  lambda_ramp_shape: linear
+  lambda_ramp_start_frac: 0.7
+  lr: 0.001
+  lr_schedule: warmup_cosine
+  prism_top_n: 1
+  row_sampling: hybrid_50freq_50uniform
+  steps: 20000
+  warmup_steps: 1000
+  weight_decay: 0.0001
+evaluation:
+  k: 256
+  max_eval_rows: 20000
+  max_eval_hidden: 1024
+  topk_overlap: 5
+  contribution_top_k: 8
+artifacts:
+  save_checkpoint: true
+  make_figures: false
+autoresearch:
+  score_thresholds:
+    max_dead_feature_rate: 0.5
+    max_val_logit_kl_bits_mean: 20.0

qwen3.5-2b/k256_32x_seedvar_s0/metrics.json

@@ -0,0 +1,39 @@
+{
+  "always_on_feature_rate": 6.103515625e-05,
+  "architecture": "topk",
+  "d_features": 65536,
+  "data_seed": 0,
+  "data_source": "/nfs-share/mh2274/readout_prism_workspace/data/qwen35-2b/qwen35_2b.pt",
+  "dead_feature_rate": 0.100799560546875,
+  "device": "cuda",
+  "duplicate_decoder_frac": 0.0,
+  "eval_hidden": 512,
+  "eval_rows": 20000,
+  "feature_usage_entropy": 0.8841440074929942,
+  "finite_metrics": true,
+  "identity_abs_error_max": 3.24249267578125e-05,
+  "identity_abs_error_mean": 5.0943344831466675e-06,
+  "init_seed": 0,
+  "k": 256,
+  "l0_controller_enabled": false,
+  "l0_controller_final_coeff": null,
+  "l0_controller_target": null,
+  "l0_p10": 256.0,
+  "l0_p50": 256.0,
+  "l0_p90": 256.0,
+  "mean_l0": 255.812,
+  "mean_l0_train": 254.47453352311751,
+  "rare_feature_rate": 0.44915771484375,
+  "row_centered_cosine": 0.9070281982421875,
+  "row_centered_ev": 0.823739543557167,
+  "selection_score": 3.114999217420358,
+  "top1_logit_residual_abs_mean": 0.5085803866386414,
+  "top1_logit_residual_frac_mean": 0.028850413858890533,
+  "top8_abs_contrib_coverage_mean": 0.7325428128242493,
+  "top8_positive_contrib_coverage_mean": 0.7803902626037598,
+  "train_seconds": 10821.384970664978,
+  "used_fallback_data": false,
+  "val_logit_kl_bits_mean": 0.20181342861592724,
+  "val_top1_match": 0.806640625,
+  "val_top5_overlap": 0.8230468750000036
+}

qwen3.5-2b/k256_32x_seedvar_s1/checkpoint.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0ce4816b260a25e23834472ae76b3c4a01b1bcb18a681a21f16349aa6cd9acf
+size 1076996537

qwen3.5-2b/k256_32x_seedvar_s1/config.yaml

@@ -0,0 +1,63 @@
+run:
+  name: qwen2b_d65536_k256_s1
+  seed: 1
+  init_seed: 1
+  output_dir: results/exp5_seedvar/qwen2b_d65536_k256_s1
+  device: auto
+data:
+  data_seed: 1
+  fallback: error
+  max_hidden: 2048
+  max_rows: null
+  path: /nfs-share/mh2274/readout_prism_workspace/data/qwen35-2b/qwen35_2b.pt
+  row_preprocessing: center_normalize
+  val_hidden: 512
+factorizer:
+  architecture: topk
+  d_features: 65536
+  encoder_init_scale: 0.02
+  k: 256
+training:
+  auxk_alpha: 0.0
+  auxk_dead_window: 1000000
+  auxk_k: 1024
+  batch_size: 4096
+  code_l2: 1.0e-06
+  final_lr_frac: 0.1
+  init_mode: row_seeded
+  l0_coeff: 0.0
+  l0_controller:
+    coeff_init: 0.0
+    coeff_max: 0.01
+    coeff_min: 1.0e-08
+    control_period: 10
+    ema: 0.9
+    enabled: false
+    err_clip: 0.3
+    kp: 0.2
+    target_l0: 0
+    warmup_steps: 200
+  lambda_mode: delayed_ramp
+  lambda_prism: 0.001
+  lambda_ramp_shape: linear
+  lambda_ramp_start_frac: 0.7
+  lr: 0.001
+  lr_schedule: warmup_cosine
+  prism_top_n: 1
+  row_sampling: hybrid_50freq_50uniform
+  steps: 20000
+  warmup_steps: 1000
+  weight_decay: 0.0001
+evaluation:
+  k: 256
+  max_eval_rows: 20000
+  max_eval_hidden: 1024
+  topk_overlap: 5
+  contribution_top_k: 8
+artifacts:
+  save_checkpoint: true
+  make_figures: false
+autoresearch:
+  score_thresholds:
+    max_dead_feature_rate: 0.5
+    max_val_logit_kl_bits_mean: 20.0

qwen3.5-2b/k256_32x_seedvar_s1/metrics.json

@@ -0,0 +1,39 @@
+{
+  "always_on_feature_rate": 4.57763671875e-05,
+  "architecture": "topk",
+  "d_features": 65536,
+  "data_seed": 1,
+  "data_source": "/nfs-share/mh2274/readout_prism_workspace/data/qwen35-2b/qwen35_2b.pt",
+  "dead_feature_rate": 0.1090087890625,
+  "device": "cuda",
+  "duplicate_decoder_frac": 0.0,
+  "eval_hidden": 512,
+  "eval_rows": 20000,
+  "feature_usage_entropy": 0.8831385530801812,
+  "finite_metrics": true,
+  "identity_abs_error_max": 3.62396240234375e-05,
+  "identity_abs_error_mean": 5.282461643218994e-06,
+  "init_seed": 1,
+  "k": 256,
+  "l0_controller_enabled": false,
+  "l0_controller_final_coeff": null,
+  "l0_controller_target": null,
+  "l0_p10": 256.0,
+  "l0_p50": 256.0,
+  "l0_p90": 256.0,
+  "mean_l0": 255.8083,
+  "mean_l0_train": 254.36040351250313,
+  "rare_feature_rate": 0.45306396484375,
+  "row_centered_cosine": 0.9065490961074829,
+  "row_centered_ev": 0.8228987753391266,
+  "selection_score": 3.135100835964664,
+  "top1_logit_residual_abs_mean": 0.5029797554016113,
+  "top1_logit_residual_frac_mean": 0.028145218268036842,
+  "top8_abs_contrib_coverage_mean": 0.748420000076294,
+  "top8_positive_contrib_coverage_mean": 0.7912746667861938,
+  "train_seconds": 11315.187881469727,
+  "used_fallback_data": false,
+  "val_logit_kl_bits_mean": 0.2218228146240548,
+  "val_top1_match": 0.822265625,
+  "val_top5_overlap": 0.8230468750000035
+}

qwen3.5-2b/k256_32x_seedvar_s2/checkpoint.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c30f537c387c66f5a92fa5321aef38278d69ede42b9f78254e288160735d7a57
+size 1076996537

qwen3.5-2b/k256_32x_seedvar_s2/config.yaml

@@ -0,0 +1,63 @@
+run:
+  name: qwen2b_d65536_k256_s2
+  seed: 2
+  init_seed: 2
+  output_dir: results/exp5_seedvar/qwen2b_d65536_k256_s2
+  device: auto
+data:
+  data_seed: 2
+  fallback: error
+  max_hidden: 2048
+  max_rows: null
+  path: /nfs-share/mh2274/readout_prism_workspace/data/qwen35-2b/qwen35_2b.pt
+  row_preprocessing: center_normalize
+  val_hidden: 512
+factorizer:
+  architecture: topk
+  d_features: 65536
+  encoder_init_scale: 0.02
+  k: 256
+training:
+  auxk_alpha: 0.0
+  auxk_dead_window: 1000000
+  auxk_k: 1024
+  batch_size: 4096
+  code_l2: 1.0e-06
+  final_lr_frac: 0.1
+  init_mode: row_seeded
+  l0_coeff: 0.0
+  l0_controller:
+    coeff_init: 0.0
+    coeff_max: 0.01
+    coeff_min: 1.0e-08
+    control_period: 10
+    ema: 0.9
+    enabled: false
+    err_clip: 0.3
+    kp: 0.2
+    target_l0: 0
+    warmup_steps: 200
+  lambda_mode: delayed_ramp
+  lambda_prism: 0.001
+  lambda_ramp_shape: linear
+  lambda_ramp_start_frac: 0.7
+  lr: 0.001
+  lr_schedule: warmup_cosine
+  prism_top_n: 1
+  row_sampling: hybrid_50freq_50uniform
+  steps: 20000
+  warmup_steps: 1000
+  weight_decay: 0.0001
+evaluation:
+  k: 256
+  max_eval_rows: 20000
+  max_eval_hidden: 1024
+  topk_overlap: 5
+  contribution_top_k: 8
+artifacts:
+  save_checkpoint: true
+  make_figures: false
+autoresearch:
+  score_thresholds:
+    max_dead_feature_rate: 0.5
+    max_val_logit_kl_bits_mean: 20.0

qwen3.5-2b/k256_32x_seedvar_s2/metrics.json

@@ -0,0 +1,39 @@
+{
+  "always_on_feature_rate": 3.0517578125e-05,
+  "architecture": "topk",
+  "d_features": 65536,
+  "data_seed": 2,
+  "data_source": "/nfs-share/mh2274/readout_prism_workspace/data/qwen35-2b/qwen35_2b.pt",
+  "dead_feature_rate": 0.0898590087890625,
+  "device": "cuda",
+  "duplicate_decoder_frac": 0.0,
+  "eval_hidden": 512,
+  "eval_rows": 20000,
+  "feature_usage_entropy": 0.8889448732525317,
+  "finite_metrics": true,
+  "identity_abs_error_max": 2.47955322265625e-05,
+  "identity_abs_error_mean": 5.09992241859436e-06,
+  "init_seed": 2,
+  "k": 256,
+  "l0_controller_enabled": false,
+  "l0_controller_final_coeff": null,
+  "l0_controller_target": null,
+  "l0_p10": 256.0,
+  "l0_p50": 256.0,
+  "l0_p90": 256.0,
+  "mean_l0": 255.82395,
+  "mean_l0_train": 254.51032439240447,
+  "rare_feature_rate": 0.431915283203125,
+  "row_centered_cosine": 0.9083909392356873,
+  "row_centered_ev": 0.8261668831110001,
+  "selection_score": 3.151505386758288,
+  "top1_logit_residual_abs_mean": 0.484148234128952,
+  "top1_logit_residual_frac_mean": 0.027414647862315178,
+  "top8_abs_contrib_coverage_mean": 0.7343347072601318,
+  "top8_positive_contrib_coverage_mean": 0.7825666069984436,
+  "train_seconds": 11153.051770210266,
+  "used_fallback_data": false,
+  "val_logit_kl_bits_mean": 0.19388584532874337,
+  "val_top1_match": 0.84765625,
+  "val_top5_overlap": 0.8429687500000036
+}