license: cc-by-nc-4.0
base_model: CohereLabs/tiny-aya-global
tags: [gguf, quantization, safety, tiny-aya, lexsi-hackathon]
Shipped ≠ Safe: quantization recipes for the Tiny Aya Global safety audit
Lexsi AI Safety & Alignment Hackathon (Track 2a), Round 1.
Audited files:
CohereLabs/tiny-aya-global-GGUF→tiny-aya-global-q4_k_m.gguf(official, reference)- Q3_K_M and Q2_K, requantized from that file with
llama-quantize --allow-requantize(llama.cpp b12359)
The low-bit weights are fully reproducible from the official file with the recipe in recipe/
(run_round1.sh); their sha256s are in the dataset repo (results/gguf_sha256.txt). They were not
re-uploaded in Round 1 because of upload bandwidth. Round 2 adds repaired checkpoints (SafeTune QReSafe /
AAQ, CircuitKIT circuit-aware mixed precision) as real weights.
Licence: inherits CC-BY-NC-4.0 and the Cohere Labs Acceptable Use Policy from Tiny Aya. Safety research use.