Checkpoints
The repository root holds the final (step 16) model. Intermediate checkpoints
are in subfolders:
from transformers import AutoModelForCausalLM repo = "mremila/z7b-wmdp-bio__5pp__unlearn-fft-npo-wd__relearn-fft"model = AutoModelForCausalLM.from_pretrained(repo, subfolder="checkpoint-8")
Table with columns: Step, Subfolder| Step | Subfolder |
|---|
| 0 | unlearned model: mremila/z7b-wmdp-bio__5pp__unlearn-fft-npo-wd |
| 4 | checkpoint-4 |
| 8 | checkpoint-8 |
| 12 | checkpoint-12 |
| 16 | checkpoint-16 (same weights as the root) |
Attack
Table with columns: Setting, Value| Setting | Value |
|---|
| Data | cais/wmdp-bio-forget-corpus |
| Steps | 16 (micro batch 1, sequence length 512) |
| Learning rate | 1e-5, linear schedule |
The full axolotl config is in training_config.yml.
Relearning trajectory
Zero-shot accuracy (%), lm-eval, as WMDP-Bio robust ↓ / MMLU-no-bio ↑:
Table with columns: Model, Step 0, Step 4, Step 8, Step 12, Step 16| Model | Step 0 | Step 4 | Step 8 | Step 12 | Step 16 |
|---|
| NPO | 29.38 / 55.22 | 26.73 / 36.11 | 46.31 / 55.58 | 49.08 / 56.48 | 49.65 / 56.64 |
| NPO + weight distance | 28.23 / 54.74 | 26.50 / 23.10 | 25.69 / 24.70 | 26.96 / 25.29 | 26.27 / 24.89 |
The base Zephyr-7B-β scores 55.18 on WMDP-Bio robust and 58.59 on MMLU-no-bio. In this
attack (seed 1, data seed 42) WMDP-Bio stays near chance (25%) for the
weight-distance model, but MMLU-no-bio collapses to chance as well from step 4
on: the attack breaks the model rather than being resisted. The collapse depends
on the seed: a second attack (seed 2, data seed 2) relearns normally, reaching
49.31 / 55.36 at step 16, close to plain NPO.