MdSadiqMd's picture
Super-squash branch 'main' using huggingface_hub
ff811bb
|
Raw History Blame Contribute Delete
3.37 kB
metadata
pretty_name: 'Circuit steering on Tiny Aya: prompts and facts'
language:
  - en
  - hi
  - ar
  - fr
  - sw
tags:
  - mechanistic-interpretability
  - ai-safety
  - refusal
  - knowledge-editing
configs:
  - config_name: refusal_prompts
    default: true
    data_files:
      - split: cyber
        path: b2/pool_cyber.jsonl
      - split: theft_fraud
        path: b2/pool_theft_fraud.jsonl
      - split: weapons_violence
        path: b2/pool_weapons_violence.jsonl
  - config_name: benign_prompts
    data_files: a22/benign.jsonl
  - config_name: xstest
    data_files: b2/xstest_all.jsonl
  - config_name: fact_prompts
    data_files:
      - split: en
        path: b1/prompts_en.jsonl
      - split: hi
        path: b1/prompts_hi.jsonl
      - split: ar
        path: b1/prompts_ar.jsonl
      - split: fr
        path: b1/prompts_fr.jsonl
      - split: sw
        path: b1/prompts_sw.jsonl
  - config_name: false_facts
    data_files: b1/false_facts.jsonl

Circuit steering on Tiny Aya: prompts and facts

The prompts, facts and splits used to find circuits in CohereLabs/tiny-aya-global, change the model, and test it, in English, Hindi, Arabic, French and Swahili. Part of the Lexsi hackathon submission Circuit-Steering (Track 2b).

This repo contains harmful requests taken from public benchmarks (the requests only). No reply to a harmful request is here.

The tables in the viewer

config splits content
refusal_prompts cyber, theft_fraud, weapons_violence harmful requests (HarmBench, MultiJail) and harmless look-alikes (XSTest) per harm category, in five languages
benign_prompts train harmless prompts used to train and tune the adapters, with the base model's replies
xstest train the 450 XSTest prompts (250 safe, 200 unsafe), English
fact_prompts en, hi, ar, fr, sw 1,502 Wikidata fact questions per language
false_facts train 180 facts the model gets wrong, with its wrong answer

In refusal_prompts, the column split says harmful or benign, and source names the benchmark. translated_by_model lists the languages Tiny Aya translated itself; bt_<lang> is the translation put back into English and bt_f1_<lang> its word overlap with the original.

from datasets import load_dataset

cyber = load_dataset("MdSadiqMd/circuit-steering-data", "refusal_prompts", split="cyber")

The other files

path content
a22/split.json which prompt ids are train, dev and test for the adapters
a23/, a24/ plain text used to measure and to protect fluency
b1/facts_pool.jsonl, b1/model_eval_<lang>.jsonl, b1/ripple_en.jsonl the fact pool, the base model's answer to every fact, related facts
b1/eap_<lang>.csv, b2/eap_<lang>.csv, b1/eap_known_<lang>.csv, b2/eap_<lang>_matched.csv clean and corrupted prompt pairs used to find the circuits (CSV; the viewer shows one file format per repo, so these are not in it)
b2/baseline_*.jsonl whether the base model refused each prompt (no reply text)
bystanders/ prompt pairs for the other circuits checked for side effects
*.py, MANIFEST.json the scripts that built these files, and file hashes

Sources: HarmBench, MultiJail, XSTest and Wikidata; each keeps its own licence. Translations and replies come from CohereLabs/tiny-aya-global (CC-BY-NC 4.0).