cross-model-lora-prediction-3b / ROUND8_REPORT.md
CK0607's picture
Round 8 gap-fill artifact: ROUND8_REPORT.md
d1e91f3 verified
|
raw
history blame
3.07 kB

Round 8 — Fill R6 Gaps

R8 fills three gaps R6 left open: K-sweep at N=24, per-tensor and Procrustes methods, and a per-task breakdown.

1A. K-sweep at N=24

K aggregate gap_recovered ± std
2 0.032 ± 0.251
4 0.069 ± 0.210
6 0.083 ± 0.136
8 0.129 ± 0.156
12 0.131 ± 0.118
16 0.129 ± 0.097
20 0.130 ± 0.089
24 0.126 ± 0.089

No K beats R6 global_ridge at N=24 (0.135); K=12 is best among Top-K rows at 0.131.

1B. Per-tensor methods

method N=12 N=16 N=24
mean 0.077 ± 0.071 0.077 ± 0.071 0.083 ± 0.072
global_ridge 0.126 ± 0.128 0.137 ± 0.126 0.135 ± 0.104
topk8_global_ridge 0.117 ± 0.109 0.124 ± 0.101 0.121 ± 0.124
pertensor_ridge 0.115 ± 0.109 0.121 ± 0.093 0.126 ± 0.109
procrustes 0.105 ± 0.082 0.104 ± 0.086 0.100 ± 0.074
pertensor_pca 0.100 ± 0.076 0.100 ± 0.080 0.114 ± 0.103

1C. Per-task paper table

Domain Task base_Y mean_N16 best_R6_N16 best_R8_new_N16 best_learned_N16 oracle gap_recovered winner
math gsm_hard 0.063 0.066±0.005 0.066±0.005 (mean) 0.070±0.007 (pertensor_pca) 0.070±0.007 (pertensor_pca) 0.150 0.077±0.077 R8:pertensor_pca
math gsm8k_test_500 0.080 0.102±0.002 0.102±0.002 (mean) 0.106±0.002 (pertensor_pca) 0.106±0.002 (pertensor_pca) 0.293 0.120±0.009 R8:pertensor_pca
code mbpp_test_held 0.230 0.240±0.000 0.257±0.006 (global_ridge) 0.250±0.000 (pertensor_ridge) 0.257±0.006 (global_ridge) 0.320 0.296±0.064 R6:global_ridge
code mbpp_plus 0.217 0.212±0.002 0.270±0.003 (global_ridge) 0.266±0.002 (pertensor_ridge) 0.270±0.003 (global_ridge) 0.450 0.229±0.014 R6:global_ridge
science openbookqa_test 0.710 0.754±0.002 0.754±0.002 (mean) 0.756±0.019 (pertensor_pca) 0.756±0.019 (pertensor_pca) 0.983 0.167±0.070 R8:pertensor_pca

Headline

Best aggregate (N, method) cell across R6 ∪ R8: R6 N=16 method=global_ridge gap_recovered=0.137 ± 0.126. Best individual task/seed cell: R6 task=mbpp_test_held N=4 seed=1 method=global_ridge gap_recovered=0.556 acc=0.280. Per-task winners over all evaluated R6 ∪ R8 cells: gsm_hard: R6:topk8_global_ridge N=8 seed=0; gsm8k_test_500: R6:mean N=12 seed=1; mbpp_test_held: R6:global_ridge N=4 seed=1; mbpp_plus: R6:topk8_global_ridge N=24 seed=0; openbookqa_test: R8:pertensor_pca N=12 seed=2.

Honest failure modes

  • No surrogate was used: every R8 accuracy comes from the same model.generate(do_sample=False, num_beams=1) path as R6.
  • Degenerate cells (gap_recovered ≤ -0.25): gsm_hard:topk2_global_ridge:N24:seed0=-0.31
  • Cross-domain/locality failures at K<5: none on aggregate, though individual tasks vary.
  • Seed instability (std > 0.15): none for 1B aggregate method×N cells.
  • Code-task evaluation remains the R4/R6 cheap string/span proxy rather than sandboxed unit tests.