| --- |
| license: cc-by-nc-4.0 |
| base_model: |
| - CohereLabs/tiny-aya-earth |
| tags: [safety, red-teaming, interpretability, abliteration, activation-steering, lexsi-hackathon] |
| --- |
| # Refusal is not Recognition: repair kit for Tiny Aya |
|
|
| Small artifacts from **Refusal is not Recognition** (Lexsi AI Safety & Alignment Hackathon, Track 2). These are not full weights. They are the vectors and probes that reproduce the monitors and the recognition-gated repair on top of the official Tiny Aya checkpoints. |
|
|
| | File | What | |
| |---|---| |
| | `directions.pt` | Earth: SafeTune refusal direction (layer 24), CAST condition vector, layer (16) and threshold | |
| | `directions_global.pt` | Global: SafeTune refusal direction (layer 24) | |
| | `calib.npz` | layer-16 recognition probe (`w`, `b`, raw residual space) and calibration scores per wrapper | |
| | `repair_meta.json` | CASTModel repair: steering alpha, plain/mixed gate thresholds, alpha sweep | |
| | `jlens_earth_100.pt`, `jlens_earth_300.pt` | Jacobian lens (anthropics/jacobian-lens) fitted on Earth, on 100 and 300 wikitext prompts | |
|
|
| Abliterated weights are deliberately **not** released; the refusal direction plus SafeTune `orthogonalize_weights` reproduces them for research. |
|
|
| Code, notebooks and full write-up: https://github.com/saiadith/refusal-vs-recognition |
|
|
