Evaluation-only card: refusal-direction steering on Tiny Aya Global

This repository contains no model weights. It points to the base model CohereLabs/tiny-aya-global at revision af89d219b53ed9b13b8a4645f9c8028973510324 and records the inference-time steering configuration used in the Lexsi Track 2 Stage 1 evaluation. The weights are not changed: SafeTune adds a direction to the output of block 24 during generation.

  • Steering: SafeTune RefusalDirectionModel(mode="steer", layers=[24], strength=15.0789); ฮฑ = 0.5 ร— direction norm 30.1578; direction SHA-256 a29d231a6e7a4734a113d66845e0849cdf25326828e0d4e6d21b26737bc246fe.
  • Refusal vector: The refusal vector is not published: only its layer, norm, strength and SHA-256 are, so the result can be checked against a re-extraction.
  • Results and logs: see the dataset repository KG306-v/tiny-aya-refusal-steering-eval. Headline numbers there; this card makes no further claim.
  • Scope: Scores come from automated judge and phrase-based labels on a small fixed prompt set; this is not a semantic safety audit and does not establish safety in other languages, prompts or attacks.
  • Licence of the base model: CC-BY-NC-4.0 (Cohere Labs).
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for KG306-v/tiny-aya-refusal-steering-config

Finetuned
(171)
this model

Collection including KG306-v/tiny-aya-refusal-steering-config