CalibratingAutorater/judgelm-preference-distributions

JudgeLM Persona Preference Distributions

Multi-annotator pairwise preference labels for training and evaluating probabilistic LLM autoraters, from the paper Judging with Confidence: Calibrating Autoraters to Preference Distributions (EMNLP 2026 Findings).

Each item is a response pair (A, B) from the JudgeLM corpus, judged by a teacher LLM under multiple sampled personas. Aggregating the persona votes yields an empirical preference distribution — a soft target p_b_over_a = Pr[B ≻ A] — instead of a single hard label, so an autorater can be trained to output a calibrated probability rather than a discrete verdict.

We release only the preference labels, not the underlying JudgeLM prompts/responses (those carry upstream licenses). The large train/RL splits are keyed by JudgeLM's review_id; run reconstruct.py to merge the original text back in (see below). The small evaluation splits are self-contained.

Splits

Config	Source	Content	Items	Votes	Votes / item	Role
gemini_sft_5k	Gemini-2.5-Flash	lean (labels)	4,240	42,321	up to 10	Dense calibration (SFT)
gemini_sft_5k_cot	Gemini-2.5-Flash	lean + teacher CoT	4,240	42,321	up to 10	Dense calibration with reasoning
gemini_rl_50k	Gemini-2.5-Flash	lean (labels)	49,999	49,999	1	Sparse calibration (RL)
gemini_eval_1k	Gemini-2.5-Flash	self-contained	997	9,970	10	Evaluation
gpt5mini_eval_1k	GPT-5-mini	self-contained	997	9,970	10	Evaluation (source-transfer)

The three core Gemini splits total 102,290 votes over 55,236 items (≈110K teacher queries were issued; votes the teacher failed to return in parsable form were dropped). sft_5k_cot is the same SFT items with teacher chain-of-thought (its votes overlap sft_5k). The GPT-5-mini split is an independent preference source for testing source-independence and cross-distribution transfer.

Data fields

Common to all splits: question_id, response_a_model, response_b_model, annotator_model, num_annotations, votes, p_b_over_a, raw_responses.

Field	Type	Description
question_id	int	JudgeLM question id.
review_id	string	(lean splits) JudgeLM review id — the join key for reconstruction.
prompt	string	(self-contained splits) the user instruction.
response_a, response_b	string	(self-contained splits) the two candidate responses.
response_a_model, response_b_model	string	Model that produced each response.
annotator_model	string	Teacher model: gemini-2.5-flash or gpt-5-mini.
num_annotations	int	Number of parsed votes (1 for RL, up to 10 otherwise).
votes	list[dict]	One record per vote: {"vote": "A"|"B", "persona_id": int, "was_swapped": bool}. Votes are in canonical A/B orientation; was_swapped marks items the teacher saw in flipped order. persona_id indexes the annotation persona (paper, Appendix E).
p_b_over_a	float	Training target p̂(x) = Pr[B ≻ A] = fraction of votes for B.
raw_responses	list[string]	Raw teacher outputs (e.g., <winner>B</winner>), in prompt (possibly swapped) orientation.
chain_of_thought	string	(sft_5k_cot only) teacher reasoning trace.

Loading

from datasets import load_dataset

# Labels are usable directly (no reconstruction needed):
rl  = load_dataset("CalibratingAutorater/judgelm-preference-distributions", "gemini_rl_50k", split="train")
y   = rl[0]["p_b_over_a"]                # soft target Pr[B > A]

# Evaluation splits include the text:
ev  = load_dataset("CalibratingAutorater/judgelm-preference-distributions", "gemini_eval_1k", split="train")

Reconstructing prompts/responses for the train/RL splits

The lean splits omit JudgeLM text. To merge it back, join on review_id against JudgeLM-100K:

pip install -U huggingface_hub
python reconstruct.py                       # downloads BAAI/JudgeLM and writes *_full.jsonl
# or: python reconstruct.py --judgelm /path/to/judgelm_train_100k.jsonl

This produces sft_5k_full.jsonl, sft_5k_cot_full.jsonl, rl_50k_full.jsonl with prompt, response_a, response_b added.

Annotation process

The teacher LLM is prompted to pick the better of two responses. To capture a diverse population, each vote conditions on a randomly sampled persona (temperature 1.0). Seven personas are used — Factual & Technical Specialist, Safety & Privacy Advocate, Empathetic Communicator, Concision Champion, Nuance & Depth Seeker, Creative & Arts Critic, and Everyday User (sampled at 4× weight to reflect the non-expert-dominated end-user population). Response order is swapped on a subset of queries (was_swapped) to mitigate positional bias; votes are normalized back to canonical A/B orientation. The empirical preference distribution for a pair is the mean of its votes.

The persona-simulated distribution is a scalable proxy for diverse human preference; it is not a real demographic population.

Source data

Prompts and response pairs are from the JudgeLM corpus (Zhu et al., 2025) — BAAI/JudgeLM on the Hugging Face Hub — which aggregates instruction-following prompts (e.g., Alpaca-GPT4, Dolly-15K) with responses from a range of open-source LLMs. Only the preference labels here are new.

Licensing

Released under Apache-2.0, consistent with JudgeLM. Labels are generated by Gemini-2.5-Flash and GPT-5-mini; use should comply with the respective providers' terms. Intended for research on autorater calibration, preference modeling, and evaluation.