tyb343/rounded-off-bank-sft
Updated • 54
Fine-tuning data and English+Hindi eval sets for Rounded Off: do safety repairs survive 4-bit? Lexsi hackathon.
Note Fine-tuning chats: Banking77 questions, Adaption answers, EN + Hindi.
Note All eval prompts, EN + HI: HarmBench, XSTest, bank red-team/borderline, bank task.
Note Banking77 seed queries (train + held-out test).
Note English public safety prompts (HarmBench, XSTest).
Note AdvBench prompts used for SafeGrad refusal pairs.
Note Bank fraud red-team (should refuse) and borderline (should answer) prompts, EN + HI. The test behind the headline.
Note BankDesk simulator: scenario cards, SFT demos, DPO pairs, GRPO states
Note Scam-victim and genuine-urgent requests to a bank assistant, EN + HI