Evaluation scope
Evaluation used text-only input on the frozen 3,668-item WMDP C set. This does not evaluate image or broader multimodal safety. The reported C result is from the frozen scoring run; this exported HF checkpoint was separately compared against the same edited full wrapper on CPU, tensor by tensor.
For base accuracy A0 = base_correct / n and edited accuracy Au = edited_correct / n, the project reports C = (A0 - Au) / (A0 - 0.25). The exact result is shown as a reduced fraction as well as a percentage.
Table with columns: WMDP C items, Base correct, Edited correct, Exact C, C| WMDP C items | Base correct | Edited correct | Exact C | C |
|---|
| 3668 | 2173 | 1211 | 481/628 | 76.5924% |
The separate text-only S evaluation is complete: raw S 96.321895%; capped S 96.321895%. See the text S results at the frozen project commit. This does not evaluate image or broader multimodal safety.
Source and edit
- Full multimodal base:
OpenGVLab/InternVL3-8B
- Frozen base revision:
853e3a797a661694b1b8ece0cb72dc2b23e3dac9
- Wrapper class:
InternVLChatModel
- Sparse update SHA-256:
1cb064d13f6370babc154197480a06705c38ebaf26a3baae8a89850419e287e8 (678,955,986 bytes)
- Changed weights: language-model MLP
down_proj matrices in layers 4, 5, 6, 7, 9.
This is a sparse GROM-family edit. The frozen update file contains the listed matrices; the contribution of a separately recorded NPO pipeline step is not established by the final overlay, so no NPO-specific effect is claimed.
Loading
Use the complete model through Transformers; no training code or overlay application is needed at load time.
import torchfrom transformers import AutoModel, AutoTokenizerrepo = "HoshiuZ/InternVL3-8B-Unlearned"tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True, use_fast=False)model = AutoModel.from_pretrained( repo, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)
License and attribution
The source checkpoint carries the Apache-2.0 model license metadata. The upstream InternVL project states that its project/code is under MIT and that the Qwen2.5 component is under Apache-2.0; see LICENSES.md, the source model card, and OpenGVLab/InternVL. The pinned remote model code is retained from the stated source revision.
Source code
Research and implementation source: HoshiuZ/multimodal-llm-safety.