GitHub | Model | Dataset | Blog

Introduction

• ReasonLite is an ultra-lightweight math reasoning model. With only 0.6B parameters, it leverages high-quality data distillation to achieve performance comparable to models over 10× its size, such as Qwen3-8B, reaching 75.2 on AIME24 and extending the scaling law of small models.
• The project is fully open-source, including model weights, training scripts, training data, and the data synthesis + filtering pipeline.

Model

The model is trained in two progressive distillation stages. First, short-CoT data is used to distill Qwen3-0.6B into ReasonLite-0.6B-Turbo, improving AIME24 accuracy from 11.0 → 57.1. Then, long-CoT data is used to obtain ReasonLite-0.6B, further boosting accuracy to 75.2.

Evaluation Results: We evaluate model performance on math reasoning tasks.

• avg@16: The average accuracy over 16 independently generated answers.
• pass@8: The probability that at least one correct answer appears among 8 generated samples.

Model	Parameters	AMC23 avg@16	AMC23 pass@8	AIME25 avg@16	AIME25 pass@8	AIME24 avg@16	AIME24 pass@8
Qwen2.5-14B	14B	58.3	82.3	12.3	32.3	12.7	32.4
Deepseek-qwen-14B	14B	93.9	98.7	50.2	71.0	65.0	83.0
Qwen3-0.6B	0.6B	52.7	85.0	16.0	33.0	11.0	31.5
Qwen3-1.7B	1.7B	83.4	96.3	36.0	55.1	47.3	73.9
Qwen3-4B	4B	96.1	100	63.5	85.4	72.7	85.1
Qwen3-8B	8B	94.8	100	68.3	84.2	74.6	85.0
Qwen3-14B	14B	98.6	98.7	71.5	84.1	78.3	88.4
DeepscaleR-1.5B	1.5B	83.8	95.0	29.0	48.9	40.4	69.0
POLARIS-1.7B-Preview	1.7B	92.2	97.4	52.3	80.2	65.0	76.7
OpenMath-Nemotron-1.5B	1.5B	88.8	96.7	39.8	65.8	61.5	81.3
ReasonLite-0.6B-Turbo	0.6B	81.6	99.3	42.7	69.2	57.1	79.6
ReasonLite-0.6B	0.6B	95.2	100	62.9	84.1	75.2	90.2

Model Link

Model	Description	AIME24	Link
amd/ReasonLite-0.6B-Turbo	Short CoT balancing performance and efficiency	57.1	🤗 HuggingFace
amd/ReasonLite-0.6B	Long CoT for high performance	75.2	🤗 HuggingFace

Dataset

• A total of 343K math problems originated from Polaris and OpenMathReasoning.
• Using the GPT-OSS model as the teacher, 9.1M model-generated raw answers were collected under both medium and high reasoning modes.
• Pseudo-labels are created by majority voting over the model outputs.
• Finally, 6.1M samples were retained, including 4.3M medium-level and 1.8M high-level reasoning data.

Dataset Link

Dataset	Description	Size	Link
amd/ReasonLite-Dataset	Short CoT	4.3M	🤗 HuggingFace
amd/ReasonLite-Dataset	Long CoT	1.8M	🤗 HuggingFace

Setup Environment

Docker:

# Recommended ROCm Docker
rocm/vllm-dev:open-mi300-08052025

# Recommended CUDA Docker (latest vLLM)
docker.io/vllm/vllm-openai:latest

Install Python packages for data synthesis and evaluation:

pip install -r requirements.txt

Model Training

This project uses open-r1 as the training codebase, but you can use any other training framework as well. Below are example launch scripts and YAML configurations used in our experiments.

# Stage 1 SFT

# Example training script:
train/stage1.sh

# Example YAML config:
train/config_stage1.yaml


# Stage 2 SFT

# Example training script:
train/stage2.sh

# Example YAML config:
train/config_stage2.yaml

Model Evaluation

Example for evaluating ReasonLite-0.6B on AIME24. The evaluation scripts are based on DeepMath.

cd eval
python3 start.py -c config/eval.yaml

Data Generation Pipeline

You will need to edit the configuration file config/oss.yaml, or create a new one based on your own needs. Then you may run through the full pipeline as follows.

Start vLLM Server

python3 vllm_start.py -c config/oss.yaml

Synthetic Data

Generate model answers

python3 infer.py -c config/oss.yaml -m infer

Pseudo-Labels via Voting (Optional)

Obtain pseudo-labels through majority voting

python3 infer.py -c config/oss.yaml -m vote

Judging Answer Correctness

Judge correctness using provided labels

python3 infer.py -c config/oss.yaml -m judge

Judge correctness using pseudo-labels (requires running vote first)

python3 infer.py -c config/oss.yaml -m judge_vote

Filtering and Converting to Training Format

Filter out incorrect solutions and convert the correct ones to training format. Specify the path to the judged data:

python3 utils/saving_to_training_format.py -d path/to/judged/data.jsonl

Directory Structure for Data Storage

datas/
└── <experiment_name>/
    ├── info.jsonl           # input prompts
    ├── answer_origin/       # raw generations
    │   └── <timestamp>/
    │       └── 0_1.jsonl
    ├── answer_judge/        # judged generations
    │   └── <timestamp>/
    │       └── 0_1.jsonl
    ├── vote/                # majority votes per prompt
    │   └── <timestamp>/
    │       └── 0_1.jsonl
    └── answer_judge_vote/   # judged using vote labels
        └── <timestamp>/
            └── 0_1.jsonl

JSONL Examples

info.jsonl

The files info.jsonl and vote/<timestamp>/0_1.jsonl follow this format:

{
    "prompt": "Solve the equation (x^2 - x - 2 = 0).",
    "expected_answer": "5",
    "index": "pol:0",
    "vote": {"x=2,-1": 5, "x=2": 1}  // Only in vote/<timestamp>/0_1.jsonl
}

• prompt contains the math problem
• expected_answer is the ground-truth label
• index is the global ID using the format "dataset_name:index"
• vote shows the voting results (only in vote files)

Intermediate Result JSONL

Files under answer_origin/<timestamp>/0_1.jsonl, answer_judge/<timestamp>/0_1.jsonl, answer_judge_vote/<timestamp>/0_1.jsonl follow this format:

{
    "info": "<input info from info.jsonl>",
    "index": "pol:0_3",
    "model_input": "<full input ...>",
    "model_output": "<full output with input prepended ...>",
    "prompt": "The front tires of a car wear out after 25,000 km, ...",
    "answer": "<model output ...>",
    "judge": true  // Only in answer_judge and answer_judge_vote
}

• Files under answer_origin contain raw model trajectories
• answer_judge and answer_judge_vote add the boolean judge flag indicating correctness

Citation

@misc{reasonlite2025,
  title    = {ReasonLite: An Ultra-Lightweight 0.6B Reasoning Model},
  author   = {An, Zihao and Chen, Chushi and Liu, Ziqiong and Li, Dong and Barsoum, Emad},
  year     = {2025},
  url      = {https://github.com/AMD-AGI/ReasonLite},
  note     = {Open-source project}
}