uzainahmed/tiny-aya-global-refusal-direction-multi

Refusal direction for Tiny Aya Global (multi contrast set)

A single unit vector in the residual stream of CohereLabs/tiny-aya-global that mediates refusal, extracted with SafeTune's extract_refusal_direction (Arditi et al., 2024) from 64 harmful and 64 harmless prompts (multi = English-only or mixed-language contrast set), layer 24. No weights were changed. Adding strength × direction to every decoder block's output at inference strengthens refusal; projecting it out removes refusal (the ablation control in the write-up).

Files: refusal_direction.safetensors (key direction, fp32, shape [hidden]), steering_config.json (layer, pooling, contrast-set SHA-256 516a30fa1eaf75246c8280aca4706a942300976fbc251a77f3d831937e563e9e, created 2026-11-13T10:53:20+00:00).

Use

from safetensors.torch import load_file
from safetune.steer import RefusalDirectionModel
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-global", dtype="float16", device_map="cuda")
tok = AutoTokenizer.from_pretrained("CohereLabs/tiny-aya-global")
direction = load_file("refusal_direction.safetensors")["direction"]
with RefusalDirectionModel(model, direction=direction, mode="steer", strength=1.0):
    ...  # model.generate(...) as usual

The strength we selected on the dev split, and the per-language before/after numbers, are in the write-up linked from the GitHub repo. This is a hackathon artifact: evaluated with an LLM judge, not human labels.