--- base_model: CohereLabs/tiny-aya-base library_name: peft license: cc-by-nc-4.0 tags: [lora, safety, refusal, agents, tool-use, interpretability, lexsi-hackathon] --- # tiny-aya-refusal LoRA adapter on `CohereLabs/tiny-aya-base` from **Refusal Is Keyed to Context, Not Content** (Lexsi AI Safety Hackathon, Track 2). Plain refusal-first chat-safety SFT (Step 1 baseline). - **Write-up & full results:** https://github.com/Lakshay768/Lexsi-Hackathon - **Recipe:** LoRA r16 a32 on all linear layers; see the write-up for data, seed and metrics. - **Load with the included `chat_template.jinja`.** ## Intended use & limits Research adapter for agent-safety / interpretability. Single-turn agent templates, English evaluation. Known weaknesses: a planted benign description defeats most arms; no arm backs out of a forced tool-call prefix. ## Safety Only adapters that act on harmful agent requests no more often than their chat-safety start are released; safety-degraded models are withheld.