# THUDM/LongBench

LongBench v2 and LongBench (ACL 25'&24')

Repository: https://github.com/THUDM/LongBench
Canonical: https://ross.abutalabs.com/products/longbench
Homepage: https://longbench2.github.io
Language: Python
License: MIT
License Family: permissive
Topics: benchmark, llm, longtext, long-context
Last push: 2025-01-15T02:04:04+00:00

## Health v2 (maintenance only)
Score: 29/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 1, release rhythm 35, longevity 80
- inputs: {"age_days": 1131, "days_push": 596, "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 1229, forks 138 (observed 2026-08-28T04:04:03.607696+00:00)

## What it is
LongBench is a benchmark suite (v1 and v2) for evaluating large language models on long-context understanding and reasoning tasks, with contexts from 8k to 2M words across six task categories. It includes 503 challenging multiple-choice questions in v2, plus datasets, evaluation code, and a public leaderboard.

## Use cases
- evaluate llm long-context understanding
- benchmark models on long document qa
- compare llms on 128k+ token contexts
- test long-context reasoning ability of a model
- measure effect of inference-time compute on long-context tasks
- find leaderboard results for long-context benchmarks

## When to choose
- you need a standardized, reliable benchmark for long-context LLM evaluation
- you want multiple-choice scoring for reproducible comparisons
- you need contexts spanning 8k to 2M words across diverse task types

## When to avoid
- you need open-ended generation quality evaluation rather than multiple-choice accuracy
- you need short-context or general-purpose benchmarks
- you need a training dataset rather than an evaluation benchmark

## Facets
- artifact type: dataset
- maturity: active
- function: benchmarking, llm-inference, nlp
- domain: large-language-models, machine-learning
- platform: python
- tags: long-context, evaluation, llm-benchmark, multiple-choice, leaderboard, natural-language-processing, algorithms

## Member repositories
- THUDM/LongBench (main) score 29

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:03.607696+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 2026-08-30T06:15:05.508657+00:00, confidence not recorded.
  - readme: https://github.com/THUDM/LongBench (fetched 2026-08-28T04:04:03.607696+00:00, sha 3333e97e2945)
  - homepage: https://longbench2.github.io (fetched 2026-08-29T12:22:56.041719+00:00, sha a76881d23555)
- Data as of 2026-08-30T08:39:29.467469+00:00.