Tiny Aya Global: multilingual refusal-steering vectors
This repo holds steering vectors, not new weights. They're applied at inference time to the unchanged CohereLabs/tiny-aya-global (3.35B). They come from a Lexsi Labs AI Safety Hackathon entry (Track 2, problems 2a.4 + 2b.8).
- Write-up: https://github.com/saihariG/tiny-aya-multilingual-safety/blob/main/ARTICLE.md
- Code: https://github.com/saihariG/tiny-aya-multilingual-safety
- Eval sets, per-row verdicts, metrics: https://huggingface.co/datasets/saiharig/tiny-aya-msafety
What's in vectors.pt
| key | value |
|---|---|
direction |
English refusal direction: difference of means (harmful − safe), last prompt token, output of decoder block 24 of 36 (2048-d) |
layer |
24 |
alpha |
27.06 (= ‖Δμ‖ at layer 24) |
gate_en |
projection gate fit on English prompts: layer 20, fires when score > 0.356, held-out F1 0.977 |
gate_ml |
projection gate fit on all languages: layer 21, fires when score > 0.644, held-out F1 0.953 |
train_rows |
3,563 prompts used to fit the direction and gates |
train_rows_excluded_overlap |
1,595 rows dropped because a translation of the same source prompt is in the eval set |
How to apply it
The intervention adds alpha_scale · alpha · unit(direction) to the output of block 24 at every
position. It's the same thing SafeTune's RefusalDirectionModel(mode="steer") does.
With SafeTune (always-on):
import torch
from huggingface_hub import hf_hub_download
from safetune.interventions.steer.refusal_direction import RefusalDirectionModel
v = torch.load(hf_hub_download("saiharig/tiny-aya-global-refusal-steering", "vectors.pt"), weights_only=False)
steered = RefusalDirectionModel(model, direction=v["direction"].float(), mode="steer",
strength=0.5 * v["alpha"], layers=[v["layer"]]) # alpha/2 recommended
out = steered.generate(**enc, max_new_tokens=256, do_sample=False)
Without SafeTune: apply_steering.py in this repo is a plain PyTorch-hook version and also
supports the gates:
python apply_steering.py --prompt "..." --alpha-scale 0.5 --gate none
Results (same model, before vs after)
These are held-out prompts: 100 MultiJail harmful + 100 XSTest-safe per language. The judge is WildGuard on back-translated responses. ASR is the harmful-answer rate on harmful prompts; over-refusal is the refusal rate on safe prompts.
| variant | ASR bn/sw/jv/hi_en % | ASR all 12 langs % | over-refusal all % |
|---|---|---|---|
| no steering | 13.2 | 7.1 | 12.4 |
| always-on, α/2 (recommended) | 8.0 | 3.8 | 23.9 |
| always-on, α | 2.5 | 1.3 | 45.3 |
| EN-gated, α | 3.5 | 1.7 | 32.7 |
| ML-gated, α | 2.8 | 1.3 | 41.2 |
Per language (ASR / over-refusal %): Swahili 9/13 → 2/19 at α/2; Bengali 11/9 → 2/21 EN-gated. These are the two languages where we confirmed the judge is reliable (77–82% precision against human review).
Limitations
- Steering raises over-refusal. At α/2 it roughly doubles.
- The gates learned topic, not intent. They over-fire on safe prompts that sound dangerous (homonyms, figurative language), so they don't remove the over-refusal cost.
- Measurement: ASR after steering is judge-measured, not human-checked. General capability under steering wasn't measured.
- Use: research use only. The base model's licence is CC-BY-NC-4.0, and these vectors follow it.