tyb343's picture
Evaluation sets
470d8ef verified
|
Raw History Blame Contribute Delete
1.56 kB
---
license: other
language: [en, hi]
tags: [safety, red-teaming, over-refusal, banking, hindi, lexsi-hackathon]
---
# Rounded Off: evaluation sets (English + Hindi)
Prompts used to evaluate every model in **Rounded Off: do safety repairs survive 4-bit?**
(code: https://github.com/tayyab415/rounded-off). Each file is JSONL with `id, lang, set, kind, prompt` (+ `reference` for task sets).
| file | rows EN / HI | source | license |
|---|---|---|---|
| harmbench_* | 150 / 148 | HarmBench standard behaviors (copyright excluded), Hindi by Adaption | MIT |
| xstest_safe_*, xstest_unsafe_* | 250+200 / 248+200 | XSTest, Hindi by Adaption | CC-BY-4.0 |
| bank_redteam_* | 250 / 250 | generated with Adaption (Invent-a-Dataset), prompts only | CC-BY-4.0 (this project) |
| bank_borderline_* | 400 / 395 | generated with Adaption, safe questions that sound risky | CC-BY-4.0 (this project) |
| bank_task_* | 308 / 155 | Banking77 test questions (4/intent) + Adaption reference answers, hi-IN localized | CC-BY-4.0 |
| rbi_qa_index_* | 150 / 150 | RBI FAQ items: ids and URLs only. RBI's text is "all rights reserved"; rebuild it with `scripts/fetch_rbi.py` and `scripts/build_data.py` | n/a |
| refusals_train.jsonl | 406 / 417 | AdvBench prompts (MIT) + Adaption refusal replies, SafeGrad safety set | MIT |
| contrast_prompts.json | 96 + 96 | AdvBench + Banking77 test, CircuitKIT circuit discovery | MIT / CC-BY-4.0 |
Harmful prompts are copied from public benchmarks or generated as red-team prompts for evaluation only; no harmful completions are included.