# GAIR-NLP/LIMO

[COLM 2025] LIMO: Less is More for Reasoning

Repository: https://github.com/GAIR-NLP/LIMO
Canonical: https://ross.abutalabs.com/products/limo
Language: Python
License Family: other
Last push: 2025-07-30T14:55:54+00:00

## Health v2 (maintenance only)
Score: 36/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 34, release rhythm 35, longevity 41
- inputs: {"age_days": 575, "days_push": 399, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1083, forks 54 (observed 2026-08-28T04:03:31.256927+00:00)

## What it is
LIMO is a research project and training framework demonstrating that large language models can achieve strong mathematical reasoning with only a small set of high-quality training samples. It releases fine-tuned Qwen2.5-32B models, curated reasoning datasets, and training/evaluation code.

## Use cases
- fine-tune an LLM for math reasoning with minimal data
- train a reasoning model on curated chain-of-thought samples
- evaluate LLMs on AIME and MATH benchmarks
- improve model generalization in mathematical problem solving
- reproduce the LIMO less-is-more reasoning results

## When to choose
- you want SOTA math reasoning via small high-quality SFT datasets
- you need training and evaluation code for reasoning fine-tuning research
- you want ready-made reasoning models and datasets on Hugging Face

## When to avoid
- you need a production inference server rather than training research code
- you lack GPU resources for 32B model fine-tuning
- you need a permissively licensed codebase - no license is specified

## Facets
- artifact type: library
- maturity: active
- function: llm-training, machine-learning, benchmarking
- domain: large-language-models, machine-learning, mathematics, artificial-intelligence
- platform: python
- tags: mathematical-reasoning, fine-tuning, sft, research, dataset, chain-of-thought, qwen, gpu, linux

## Member repositories
- GAIR-NLP/LIMO (main) score 36

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:31.256927+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 2026-08-30T06:50:12.167580+00:00, confidence not recorded.
  - readme: https://github.com/GAIR-NLP/LIMO (fetched 2026-08-28T04:03:31.256927+00:00, sha 2bac684b715f)
- Data as of 2026-08-30T08:39:29.467469+00:00.