rounded-off-tiny-aya-bank-sft-wiseft

The bank assistant after a WiSE-FT repair (halfway interpolation back to Tiny Aya Global).

Part of Safe on the benchmark, open for fraud (Lexsi Labs AI Safety & Alignment Hackathon, Nov 2026). Code and full results: https://github.com/tayyab415/rounded-off

How it was made: SafeTune WiseFTTrainer alpha 0.5 applied to rounded-off-tiny-aya-bank-sft.

Base model and license: derived from CohereLabs/tiny-aya-global (CC-BY-NC-4.0); this checkpoint carries the same non-commercial license. Built with Lexsi Labs' SafeTune (Lexsi Source Available License): this model is not for deployment (gate result below).

Intended use: research on safety drift, repair and quantization. Not a production banking assistant.

Bank release gate (article section 7.7): FAIL. It helps with fraud on 19.2% (English) / 22.0% (Hindi) of the bank red-team prompts, and 22.8% / 26.0% at 4-bit; the gate passes at 5% or less. As a BankDesk agent it sends a coached scam victim's money to the scammer in 94.8% / 83.0% of scam cases.

Downloads last month
74
Safetensors
Model size
3B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for tyb343/rounded-off-tiny-aya-bank-sft-wiseft

Finetuned
(171)
this model
Finetunes
1 model

Collection including tyb343/rounded-off-tiny-aya-bank-sft-wiseft