Mateenah/refusal-categories-results
Refusal Direction Structure by Harm Category Measurements from a mechanistic-interpretability probe asking whether refusal in instruct-tuned language models decomposes by harm category, or whether one shared direction mediates all of them. Code and full write-up: https://github.com/MateenahJAHAN/refusal-categories What this contains This repository holds measurements, not model artifacts. File Contents results.json Layer sweep, cosine-similarity matrix… See the full description on the dataset page: https://huggingface.co/datasets/Mateenah/refusal-categories-results.
| Author / organization | Mateenah |
|---|---|
| License (as declared) | mit |
| Downloads | 0 (as reported by Hugging Face) |
| Likes | 0 |
| Created | 24 October 2026 |
| Last updated at source | 24 October 2026 |
Recent changes
- 24 Oct 2026 Description: none → Refusal Direction Structure by Harm Category Measurements from a mechanistic-interpretability probe asking whether refusal in instruct-tuned language models decomposes by harm category, or whether one shared direction mediates all of them.
- 24 Oct 2026 Topics added interpretability, mechanistic-interpretability, ai-safety, refusal, evaluation
Changes detected by Silytics between refreshes of the source.
Silytics shows public metadata from Hugging Face and links to the original. Information may be incomplete or out of date. Last refreshed 24 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.