license: cc-by-nc-4.0
base_model: CohereLabs/tiny-aya-global
tags: [lora, peft, safety, multilingual]
tiny-aya-global-unalign-lora
Content note. Built for safety research in the Lexsi Labs AI Safety & Alignment Hackathon (Track 2b). This repository contains or enables harmful model behaviour. Do not deploy.
LoRA (r = 16) on tiny-aya-global, fine-tuned for 65 steps on the 100 harmful pairs only (RESTA paper recipe). Its negative is the safety vector Δ₁ used for repair: Δ₁ = aligned − un-aligned. Loaded as is, it un-aligns the model.
Prompt format: minimal Tiny Aya chat template, no system turn. Code, run log and report: https://github.com/Raj-Taware/tiny-aya-safety-drift-repair