llm judging

AlpacaEval with a length-controlled winrate estimator that neutralizes the judge's preference for longer answers.

Stanford's automatic instruction-following benchmark that compares a model's outputs to text-davinci-003 via a strong LLM judge and reports win rate.

Mercor's expert-graded eval - domain experts (doctors, lawyers, engineers) grade model responses on long-form professional tasks they would actually be paid to do.

500 challenging real-user prompts mined from Chatbot Arena and graded by a strong LLM judge for pairwise win rate.

80 two-turn open-ended questions across 8 categories, graded by GPT-4 as judge to score multi-turn dialogue quality.

800,000 step-level human labels on GPT-4 solutions to MATH problems - the canonical process-reward training/eval dataset.

2025 successor to RewardBench - harder, multi-completion (not just chosen-vs-rejected), with refreshed prompts to address contamination.

2,985 prompt-chosen-rejected triples across chat, reasoning, safety, and code - a benchmark for evaluating reward models and LLM judges.

1,024 real, long, challenging user prompts mined from WildChat conversations and judged by GPT-4 with rubric-based pairwise scoring.

NVIDIA's permissively-licensed human-annotated preference dataset with 5-axis Likert ratings - engineered to train high-quality reward models.

A self-synthesis method (and family of datasets) that elicits high-quality instructions directly from an aligned LLM using only its chat template - no seed prompts required.

Berkeley NEST's seven-way ranked preference dataset built from GPT-4 rankings over responses from a diverse model pool, used to train Starling.

OpenBMB's 64k-prompt preference dataset built with GPT-4 critiques across instruction-following, truthfulness, honesty, and helpfulness - the de facto open DPO baseline.

1M real user-chatbot conversations collected by Allen AI / UW via a free GPT proxy - a window into how real users actually prompt LLMs.

A high-quality DPO derivative of LDJnr's Capybara, with chosen/rejected pairs synthesized and rated using Argilla's distilabel pipeline.

Teknium's million-row aggregation of high-quality GPT-4-style synthetic instructions that became the de facto open SFT baseline of 2023-2024.

Evaluates pair-wise answers from RewardBench datasets

Tsinghua / OpenBMB's large-scale multi-turn dialog dataset generated by two LLMs talking to each other across structured topic taxonomies.

Microsoft's "Evol-Instruct" recipe - automatically rewriting simple instructions into harder, more diverse ones using an LLM evolver.

APEX-Agents benchmark: 480 professional services tasks across Law, Investment Banking, and Management Consulting

LMSYS's automated pipeline for distilling high-quality LLM benchmarks from crowdsourced chat data (e.g. Chatbot Arena, WildChat), producing the Arena-Hard-Auto benchmark.

LDJnr's multi-turn reasoning dataset built with the Amplify-Instruct synthesis method - short but deep conversations on a single topic.

The original community-scraped corpus of ChatGPT conversations that bootstrapped Vicuna and the entire open-instruction-tuning era.

Allen AI's flagship open SFT mixture combining new persona-driven prompts with curated public data for post-training a frontier-quality instruct model.