GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
mremila
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
mremila
Model Tree
HuggingFaceH4/zephyr-7b-beta
Input Modalities
Output Modalities
Supported Functionality
| Setting | Value |
|---|---|
| Trainer | trainers.wd.WDNPOTrainer |
| Forget / retain data | cais/wmdp-bio-forget-corpus / Salesforce/wikitext (wikitext-2) |
| Steps | 125 (micro batch 8, constant LR) |
| Learning rate | 5e-6 |
| NPO beta / retain coefficient | 0.01 / 1 |
| Weight-distance coefficient | 1e4 (selected from {0, 1e2, 1e3, 1e4, 1e5, 1e6}) |
The full axolotl config is in training_config.yml.
Zero-shot accuracy (%), lm-eval:
| Model | WMDP-Bio robust ↓ | MMLU-no-bio ↑ | L2 distance from base |
|---|---|---|---|
| Zephyr-7B-β (base) | 55.18 | 58.59 | 0 |
| NPO (5pp) | 29.38 | 55.22 | 1.94 |
| NPO + weight distance (this model) | 28.23 | 54.74 | 5.47 |
Under a full-parameter relearning attack on the WMDP-Bio forget corpus, one attack seed collapses the model (WMDP-Bio and MMLU-no-bio both near chance), while a second seed relearns WMDP-Bio to 49.31, close to plain NPO, so the weight-distance term does not give reliable tamper resistance. See mremila/z7b-wmdp-bio__5pp__unlearn-fft-npo-wd__relearn-fft.