lm-evaluation-harness
12,298A framework for few-shot evaluation of language models.
DetectionPost-generationFull HITLStableTier 2BEnvironment
Maturity tier details
Computed tier: 2B(raw score 4.64). Effective tier: 2B.
Facets:documented_in=0evaluated_on=0integration_surfaces=1paper_citations=0recent_commit=2stars=1.64
Source of truth:pipeline/maturity/score.py
Description
EleutherAI's canonical few-shot LLM evaluation framework. Implements 200+ standard benchmarks (MMLU, HellaSwag, ARC, TruthfulQA, HumanEval, GSM8K, BBH) in a unified harness. Backbone of Hugging Face's Open LLM Leaderboard.
Why it's included
Foundational LLM eval harness: cited as 'the harness' in leaderboard and model-release contexts. Differentiated from lighteval by a larger task catalog and older, broader adoption.
Addresses failure modes
- Fabrication — Generating reference to non-existent APIs, packages, or symbols.
- Logic error — Correct API use, incorrect logic.
Techniques
few-shot-evalstandard-benchmarksreference-harness