# lmarena/arena-hard-auto

Arena-Hard-Auto: An automatic LLM benchmark.

Repository: https://github.com/lmarena/arena-hard-auto
Canonical: https://ross.abutalabs.com/products/arena-hard-auto
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2025-06-21T04:44:12+00:00

## Health v2 (maintenance only)
Score: 39/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 27, release rhythm 35, longevity 73
- inputs: {"age_days": 1033, "days_push": 438, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1060, forks 159 (observed 2026-08-28T04:03:25.818452+00:00)

## What it is
Arena-Hard-Auto is an automatic benchmark for evaluating instruction-tuned LLMs using challenging real-world prompts and LLM-based judges (GPT-4.1, Gemini-2.5). It provides a cheap, fast approximation of Chatbot Arena human preference rankings with high correlation to LMArena results.

## Use cases
- benchmark my instruction-tuned LLM automatically
- estimate how a model ranks on Chatbot Arena before deployment
- compare LLMs on hard open-ended prompts
- evaluate models on creative writing tasks
- run LLM-as-judge evaluations
- measure model separability on challenging queries

## When to choose
- you need a fast, cheap proxy for human preference evaluation of chat LLMs
- you want to predict Chatbot Arena performance before release
- you need challenging real-world prompts including software engineering, math, and creative writing

## When to avoid
- you need deterministic task-specific benchmarks like MMLU or code correctness tests
- you cannot afford API calls to commercial judge models
- you need human preference data rather than LLM-judge approximations

## Facets
- artifact type: dataset
- maturity: active
- function: benchmarking, llm-inference
- domain: large-language-models, machine-learning, artificial-intelligence, testing
- platform: python, cli
- tags: llm-evaluation, llm-benchmark, chatbot-arena, model-ranking, llm-judge, leaderboard, evaluation

## Member repositories
- lmarena/arena-hard-auto (main) score 39

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:25.818452+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T06:56:57.957454+00:00, confidence not recorded.
  - readme: https://github.com/lmarena/arena-hard-auto (fetched 2026-08-28T04:03:25.818452+00:00, sha 5d04bc38063f)
- Data as of 2026-08-30T08:39:29.467469+00:00.