ConspiracyBench / eval /README.md
stationhouse's picture
Add experiment_requests/ and completed_experiments/ folders (#408)
4f9e6db
|
Raw History Blame Contribute Delete
3.86 kB

ConspiracyBench evaluation pipeline

Early, work-in-progress tooling for actually scoring a model against the corpus, as opposed to the scripts/ tools, which validate and maintain questions/*.jsonl itself.

  • judge_prompt.py — builds the prompt an LLM judge sees for one (item, model_answer) pair, composing ground_truth_answer with the item's diverges_if / does_not_diverge_if / notes guidance when present. See SCHEMA.md. For an item with pressure_turns, pass the preceding exchange as transcript=[{"role": ..., "content": ...}, ...] so the judge grades the latest response with the pushback in view. Omitting it leaves the single-turn prompt byte-for-byte unchanged.

  • score_outcomes.py — aggregates a list of judge outcomes (hold / diverge / abstain) into divergence rate (over answered items) and abstention rate (over all items). Divergence rate is n/a when every item is abstained, rather than a misleading 0%. summarize_trajectories() takes one list of per-turn outcomes per item and reports concession_rate (hold → diverge) and retreat_rate (hold → abstain) separately, plus wavered for items that diverge mid-conversation and recover by the last turn, which endpoint-only comparison hides. summarize_samples() takes k independent opening-turn outcomes per item (DEFAULT_K = 5) and reports worst_of_k, all_k_agree and pass_at_k, each as a count, rate and 95% Wilson interval, plus the pooled per-sample rates. See Repeated sampling for why pass@k is not the headline.

  • judge_cases.jsonl — a small regression/smoke set of (item, model_answer, expected_label) cases for exercising a judge implementation: clean holds, accurate extra context, false balance, direct denial, unsupported expansion, genuine uncertainty, evasive non-answers, and answer-embedded prompt injection attempting to override the expected label. Not a representative benchmark sample.

  • judge_cases_multiturn.jsonl — multi-turn companion to judge_cases.jsonl for items with pressure_turns. Each case adds a transcript (the question, earlier model answers, and the item's own pressure turns in order) and grades only the final model_answer. Covers concession after earlier holds, partial retreat into false balance, an accurate caveat that should still hold, disengagement that should abstain, recovery after an earlier concession, and rebutting a conspiracy claim by name. Kept as a separate file because the judge-reliability runners expect exactly the 15 single-turn cases in judge_cases.jsonl. test_judge_cases_multiturn.py checks every case is well formed without calling a judge.

  • completed_experiments/judge_injection/ — frozen synthetic cases and reproducible detector experiments, with labeling criteria, protocols, saved scans, and reports. The archive does not change questions/corpus.jsonl or production judge defaults.

  • test_judge_prompt.py and test_score_outcomes.py — unittest coverage for both, no dependencies beyond the standard library. Run them from the repo root with python -m unittest discover -s eval -p "test_*.py". The single-turn prompt is pinned explicitly: a silent change there rescores every existing item, so transcript=None and transcript=[] are asserted to produce the untouched prompt.

The regression cases above do not call a model. The separate completed_experiments/judge_injection/ archive documents detector studies and an optional judge-quality replay.