Tiny Aya Global β Inference-Time Refusal Steering Artifacts
Not a fine-tuned model or complete checkpoint. These are small residual-stream steering vectors extracted from and applied to frozen CohereLabs/tiny-aya-global using SafeTune.
Included artifact bundle
refusal_direction_layer18.ptβ layer-18 steering direction.cast_layer_scan_vectors.ptβ six condition vectors, layers 8/12/16/18/20/24.layer24_candidate_config.jsonβ frozen CAST development configuration.
Intended use and method
Research into static and conditional inference-time safety steering. Model weights remain frozen. Static steering used Ξ±=4 at layer 18; CAST checked a development-selected gate at layer 24 before deciding whether to steer. Conditions are not learned by modifying Tiny Aya weights.
Evaluation and limitations
On the frozen 40-Toxic benchmark, static steering changed safety-judge-flagged outputs from 2/40 to 0/40, but increased Hard-set refusal rates from 19/40 to 26/40. CAST did not improve held-out Toxic safety: 2/40 unsafe. These estimates are imprecise and classifier-based. Do not describe the vectors as a generally validated safety safeguard.
The full base model is subject to Cohere's CC-BY-NC-4.0 and acceptable-use requirements. Do not re-upload the gated base weights as if owned by this project. The project owner confirmed redistribution clearance for these artifacts on November 16, 2026; this is an owner attestation, not independent legal review or a license grant. Upstream terms and applicable attribution requirements continue to apply.
No project license is asserted in this model card. SafeTune is LSAL v1.2 source-available, not OSI open source. See the GitHub repository's LICENSE_REVIEW.md for reviewed sources and release-scope notes.
See the GitHub repository's README and REPRODUCIBILITY.md for usage and untested runtime dependencies. Fresh routing is not independently safety- or utility-validated. Model and tokenizer revisions were not recorded; exact original generation reproduction remains blocked.