mlp-surgery — restore top 10 (spec) on Qwen2.5-3B
Qwen2.5-3B-Instruct that was fine-tuned on perplexity-filtered OpenHermes 2.5 (which damaged its reasoning), then partially restored by copying back the top-10 most-damaged MLP layers from the base model. No retraining. Just weight surgery.
Method (short)
- Take the broken finetune (mlp-surgery-broken).
- Score MLP layer parameters via specificity ratio (error gradient norms / correct gradient norms).
- Copy the top-10 from base into the broken model. Save.
Eval
lm-eval, GSM8K flexible-extract 5-shot, ARC Challenge acc_norm 0-shot, no chat template, batch_size 8, single seed (2026-05-07).
| Model | GSM8K | ARC Challenge |
|---|
| Base (Qwen2.5-3B-Instruct) | 63.15% | 48.12% |
| After SFT (broken) | 61.64% | 45.22% |
| Restore top 5 | 63.00% | 45.73% |
| Restore top 15 | 63.46% | 46.50% |
| Restore top 30 | 64.29% | 48.55% |
| Restore specificity top 10 | 61.64% | 45.22% |
This model is the "Restore specificity top 10" row.
Why publish a model that didn't work?
Specificity scoring was supposed to isolate math-specifically-broken layers from generally-important ones. It didn't — restoring the top-10 specificity layers performed worse than restoring the same number of raw-norm-ranked layers. Publishing this is the negative subresult kept open. The Layer 2 case is a clean illustration: it's the top raw-score outlier (84.14 vs 12.96 for #2) but a non-outlier in specificity (1.15), correctly identifying it as a foundation layer that shouldn't be reset.
Companion models + code
Caveats
Single seed. Magnitudes are ~1pt. The "no chat template" eval style means absolute numbers are below what you'd see with chat template applied (~78% GSM8K), but relative comparisons across the same setup are meaningful.
