LIMO (Less is More for Reasoning) is a research project developed by the GAIR-NLP team and has been accepted by COLM 2025. Its core idea is to challenge the traditional view: models can achieve superior mathematical reasoning performance with less but higher-quality training data.
Purpose
LIMO is aimed at research and development on the reasoning capabilities of large language models, focusing on exploring the impact of training data quality and quantity on model reasoning performance. The project provides models, datasets, and evaluation tools for researchers and developers to use.
Main Capabilities
- Using only 817 training samples, LIMO scored 44.6 on the AIME 2025 evaluation, which is competitive compared with models trained on 800k samples.
- It demonstrates strong generalization ability; third-party evaluations show improvements on the AIME24, GPQA, and MATH-500 benchmarks.
- It provides comprehensive evaluation tools, supporting rule-based evaluation and model-based evaluation.
- It has released multiple model and dataset versions, including the latest LIMO-v2.
Models and Datasets
Both models and datasets are released through Hugging Face. The models are fine-tuned based on Qwen2.5-32B-Instruct and are compatible with frameworks such as HF Transformers, VLLM, and TensorRT-LLM.
- Models: GAIR/LIMO (the original model trained on 817 samples), etc.
- Datasets: GAIR/LIMO (the original 817 training samples), the LIMO-v2 dataset, etc.
Usage
- Environment setup: configure the training environment according to the official LLaMA-Factory documentation.
- Data preparation: obtain the LIMO-v2 dataset from Hugging Face and format it according to the LLaMA-Factory guide.
- Configuration: use the configuration files provided by the project, including hyperparameters and training settings.
- Training: multi-node training is supported, and parameters such as PET_NNODES, PET_NODE_RANK, MASTER_ADDR, and MASTER_PORT need to be set.
- Evaluation: evaluation scripts are provided, including VLLM-based inference and rule/model evaluation; for complex responses, Qwen2.5-32B-Instruct is used as the judge model. See eval/README.md for detailed instructions.
Applicable Scenarios
Suitable for large language model mathematical reasoningresearch, exploration of training data efficiency, evaluation of model generalization ability, and other scenarios.
Limitations
Multi-node training environments vary due to machine and cluster configurations, so training configurations and launch commands need to be adjusted according to the specific infrastructure. Currently, no commercial information such as fees, pricing, or platform support is provided.
The project is open-sourced under the MIT License.
