CounterSteer/sycophancy-data

Sycophancy experiment data

Data for the sycophancy steering experiments, split by what actually varies. Everything outside the model folders is shared by every model; each model folder holds the replies that model produced, which is what its probe trains on.

eval/data.json             evaluation prompts: validation + six held-out test units   (shared)
splits/*.jsonl             question-level splits, one question per split, no overlap         (shared)
probe_pool/                the TriviaQA pool and its generated wrong answers          (shared)
prompts/                   extract.jsonl and gate.jsonl, prompts with no responses    (shared)
judge_spotcheck/           blind human labels validating the belief judge             (shared)
<Model>/d_bad.jsonl        512 of the model's own sycophantic replies -> LoRA probe   (per model)
<Model>/caa_pairs.jsonl    same-question (sycophantic, correcting) reply pairs -> CAA (per model)
<Model>/stage0/            the raw pool those two are drawn from, plus judge verdicts (per model)

Models: Qwen-2.5-3B-Instruct, Qwen-2.5-7B-Instruct, Llama-3.1-8B-Instruct, Gemma-2-9B-it. 103 MB total, of which ~92 MB is stage0/.

Why the probe data cannot be shared

The over-refusal experiments pair benign prompts with a fixed, canned refusal, so one dataset serves every model. Sycophancy does not: the probe is trained on on-policy replies — the model's own agreements with a user's wrong answer, drawn from Stage 0 and labelled by a judge. A direction extracted from one model's replies is not defined for another, so d_bad.jsonl and caa_pairs.jsonl differ per model by construction.

What is shared is every prompt: the questions, their splits, the evaluation set and the two probe prompt files. prompts/extract.jsonl and prompts/gate.jsonl are byte-identical across the four models — the build checks this and refuses to continue otherwise — so they are stored once.

	rows	note
d_bad	512	per model; the LoRA probe's training set
caa_pairs	512 / 437 / 467 / 299	per model (3B / 7B / Llama / Gemma); a question contributes a pair only when the model produced both a sycophantic and a correcting reply
stage0	37,289	per model: 1 greedy + 8 sampled replies per probe question, plus judge verdicts

Evaluation splits (shared)

eval/data.json holds 7 units. validation selects the layer and alpha; the rest are held out.

unit	rows	what it measures
validation	768	256 questions × 3 user framings
test_belief	2960	740 questions × 4 framings — the headline sycophancy metric
test_pushback	300	answer, then the user pushes back with no new argument
test_feedback	1500	comment on an artifact under five user framings, with a 1–10 rating
test_feedback_free	1500	the same prompts without the rating line, judged pairwise
test_mc	1000	Anthropic multiple-choice sycophancy with a user persona
capability	128	MMLU sample as a capability guard

splits/ is the question-level layer behind validation and test_belief, plus the probe splits (train, extract, gate). Splits are disjoint by question.

Files and integrity

manifest.json lists every file with its sha256 and row count, the per-model counts and selection rules copied from each probe manifest, and the role of each shared group. build.py regenerates the folder from the experiment runs.

Each d_bad row carries the prompt, the model's reply, the judge verdict, the sampling index and the token count; caa_pairs rows carry the question with both replies.

Provenance and caveats

• Questions come from TriviaQA (probe pool) and the meg-tong/sycophancy-eval trivia subset (evaluation); the wrong answers in probe_pool/questions.jsonl were generated by Qwen3-32B and filtered, and the Stage 0 verdicts come from the same model used as a judge.
• Four of the seven evaluation units — test_pushback, test_feedback, test_mc, capability — were copied from an earlier letter-answer pilot, and test_feedback_free was assembled by hand from test_feedback. No current stage regenerates them; the rest of the data is reproducible from probe_pool/ and splits/.
• Gemma-2-9B has Stage 0 and probe data but no steering run. It is here for completeness.
• Upstream datasets keep their own licences; this folder redistributes them for reproducibility.