.gitignore

# Collected responses. These are yours, not ours -- and the safety runs contain
# real jailbreak prompts and real model completions. They must never be committed.
# "out/" is the path the READMEs pass to --out and --output-root in their
# examples, so it is ignored too; check any other path you choose with
# `git check-ignore -v <path>` before committing.
utility/output/
safety/Result_Safety/
fingerprint/met/output/
fingerprint/llmmap/output/
out/

# Downloaded benchmark files (see utility/dataset/README.md for how to get them)
utility/dataset/gpqa/gpqa_diamond.csv
utility/dataset/MedQA/test_us.jsonl
utility/dataset/LegalBench/scalr/test.tsv

# Caches: the StrongREJECT judge templates and the JBB behaviour snapshot
safety/cache/

# Your own endpoint configuration -- it names the endpoint you are auditing
endpoints.json

# Comparison tables written by --csv. They carry your endpoint label and its
# numbers, so they are ignored by default; commit one deliberately with
# `git add -f` if you mean to.
/*.csv
/results/

# NB: no bare "*.txt" rule. requirements.txt is the only .txt in the repo, and
# ignoring it would break the first step of the README. The per-run transcripts
# that run_eval.py writes live under utility/output/, already ignored above.

checkpoint_*.jsonl
__pycache__/
*.py[cod]
.env
.env.*
*.key
.DS_Store