Knows But Doesn't Refuse — models - a sunit12 Collection

sunit12 's Collections

Knows But Doesn't Refuse — data & logs

Knows But Doesn't Refuse — models

Knows But Doesn't Refuse — models

updated about 9 hours ago

Lexsi×Cohere hackathon: probe-gated refusal guard for Tiny Aya Global on romanized Hindi/Punjabi.

sunit12/tiny-aya-global-codemix-refusal-guard

Updated about 18 hours ago

Note Gate (L9) + refusal vectors + guard.py
CohereLabs/tiny-aya-global

Text Generation • 3B • Updated Sep 30 • 28.8k • • 207

Note Base model studied
CohereLabs/tiny-aya-fire

Text Generation • 3B • Updated Sep 30 • 1.44k • • 32

Note South-Asian specialist, compared
sunit12/tiny-aya-fire-codemix-refusal-guard

Updated about 2 hours ago

Note Guard re-fit on Tiny Aya Fire: usable refusal 47% -> 76%
Running

Lost in Transliteration — Explorer

🔎

Every held-out output, plain vs guarded vs CAST

Note Explorer of every held-out output
sunit12/tiny-aya-global-codemix-refusal-lora

9.24M • Updated about 2 hours ago • 31

Note Fix 2 (v4): circuit-guided LoRA, layers 10-20 — 90% usable refusal, MMLU unchanged, GGUF adapter included