Tiny Aya Fire Banking SFT (drifted baseline, research artifact)

LoRA SFT (AlignTune, r=16, α=32) of CohereLabs/tiny-aya-fire on 8,700 trilingual Bitext banking conversations, merged. This is the "before" model in a safety-drift study and is less safe than its base: refusal on banking fraud requests fell from 67% to 52% (p=0.002) and Hinglish fraud compliance rose to 40%. Use ashish-soni08/tiny-aya-fire-banking-safe instead.

Code and full results: https://github.com/Ashish-Soni08/lexsi-safe-hinglish-banking

Downloads last month
21
Safetensors
Model size
3B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ashish-soni08/tiny-aya-fire-banking-sft

Finetuned
(27)
this model

Dataset used to train ashish-soni08/tiny-aya-fire-banking-sft

Collection including ashish-soni08/tiny-aya-fire-banking-sft