| license: cc-by-nc-4.0 | |
| base_model: CohereLabs/tiny-aya-fire | |
| tags: [safety, tamper-resistance, cross-lingual, low-rank-projection] | |
| extra_gated_prompt: "For safety research. This model was adversarially fine-tuned then repaired; do not redistribute or deploy without review." | |
| # Tiny Aya Fire — Bengali-attacked, low-rank-projection repaired (k=1) | |
| `CohereLabs/tiny-aya-fire` (rev `73500c80fcb6c8bea3e830c21d5d2920249a0e1a`) fine-tune-attacked in **bn** (undefended plain SFT, harmful fraction 0.2), | |
| then repaired by removing the **top-1 singular directions** of the attack's weight update `dW = W_post - W_base` | |
| per attention/MLP projection matrix: `W_repaired = W_base + (dW - top-1 SVD component)`. | |
| Post-hoc, training-free tamper-repair (Lee 2610.00320). At k=1 this drops Bengali harm ~50%→14% while preserving | |
| MMLU — the harmful write is nearly rank-1. **Gated, for safety research only.** See the project writeup and the | |
| `tiny-aya-tamper-evals` dataset for full evaluation. | |