｜🤗 HuggingFace Model ｜
📄 arXiv ｜
🌐 中文 ｜
MedResearcher-R1 is a comprehensive training data generation and synthesis framework that tackles the challenge of domain-specific AI reasoning through knowledge-informed trajectory synthesis. Our framework provides an end-to-end solution for generating high-quality training data, consisting of three integrated components:
🧠 Knowledge Graph Construction: Our core innovation - an intelligent knowledge graph construction and QA synthesis system that transforms domain knowledge into high-quality question-answer pairs with automated reasoning path generation. This module serves as the foundation for creating domain-specific training data.
🔄 Trajectory Generation Pipeline: End-to-end trajectory synthesis and optimization system that converts QA pairs into multi-turn reasoning trajectories with tool interactions and quality filtering for model training.
📊 Evaluation Pipeline: Comprehensive model evaluation and validation framework for assessing reasoning performance across multiple benchmarks and validating the quality of synthesized training data.
These three components form a complete training data production pipeline from knowledge extraction to model training data generation and evaluation, enabling the creation of specialized reasoning models for domain-specific applications.
Features
Using our knowledge-informed trajectory synthesis framework, we developed MedResearcher-R1, a specialized reasoning model that demonstrates exceptional performance across multiple challenging benchmarks including MedBrowseComp, GAIA, and XBench-DeepSearch.
Open-Sourced Dataset
We have open-sourced a high-quality QA dataset constructed through our KnowledgeGraphConstruction module. The dataset is available at TrajectoryGenerationPipeline/qa_data/open_data.jsonl and contains:
- Complex reasoning question-answer pairs Multi-hop qa-pairs generated using our graph method
- Detailed step-by-step reasoning paths for each question, providing comprehensive problem-solving guidance
News
- [2025.8] 🎉 Our framework for generating qa and trajectory for training is officially released!
Links
Installation
MedResearcher-R1 environment
Note: This project requires Python version >= 3.10. Please ensure your environment meets this requirement.
Using venv
# create venv
python -m venv .venv
source .venv/bin/activate
# install requirements
pip install -r requirements.txt
Using conda
# create conda environment with specified Python version
conda create -n med_researcher python=3.10
# activate environment
conda activate med_researcher
# install requirements
pip install -r requirements.txt
Quick start
Train a domain-specific reasoning agent with knowledge-informed trajectory synthesis.
(1) setup environments
set -a
# fill your env in example
source env.example
set +a
(2) (optional) run graph web server and use our frontend to know how to generate qa and evaluate/filter the quality of generated qa
python KnowledgeGraphConstruction/start_web.py
Then you can use the frontend interface at http://localhost:5000. You can start with the Single QA Testing page to understand the generation process before running batch operations.
📖 For detailed feature descriptions, please refer to: features-guide.md
(3) use script to batch generate or use our provided dataset
cd KnowledgeGraphConstruction
# run batch generation - higher max-iterations will generate more complex question-answer pairs
python batch_qa_cli.py --seed-file demo_medical.csv --output ../TrajectoryGenerationPipeline/dataset/qa.jsonl --max-iterations 1
# alternatively, you can use our provided open-sourced dataset
# cp ../TrajectoryGenerationPipeline/qa_data/open_data.jsonl ../TrajectoryGenerationPipeline/dataset/qa.jsonl
(4) Launch trajectory generation and then postprocessing pipeline.
Configure the trajectory generation pipeline:
First, update TrajectoryGenerationPipeline/src/trajectory_generation/config.json with your model settings. You need to modify the following configuration parameters:
llm_config.api_key_env: Environment variable name for your API key (you still need to set up the actual environment variable). Example: API_KEYllm_config.api_base: The API base URL of your model providergeneration.model: The model name from your provider. For OpenRouter, you can check available model names at https://openrouter.ai/modelsgeneration.dataset: The dataset file in TrajectoryGenerationPipeline/qa_data directory (only accepts JSONL format). You can use our provided open_data.jsonl or generate your own dataset
Important: The read tool requires an OpenRouter API key. Either:
- Set your
OPENROUTER_API_KEY environment variable, or
- Modify the LLM client in
tools/tool_visit.py (around line 73) to use your preferred API.
cd ../TrajectoryGenerationPipeline
python src/trajectory_generation/run_reasoning.py
python src/postprocessing/pipeline.py --input_dir generation/your_model_name/your_dataset
--mode eval_filter
python src/postprocessing/pipeline.py --input_dir generation/your_model_name/your_dataset
--mode rewrite
(5) Use the rewritten_results.jsonl for training
(6) When you finish the training of your model, you can evaluate the model by creating a server for the model via vllm or sglang
pip install sglang[all]
CUDA_VISIBLE_DEVICES=0,1 python -m sglang.launch_server --model-path /path/to/your/model --port 6001 --host 0.0.0.0 --mem-fraction-static 0.95 --tp-size 2
(7) Evaluate model performance using the Evaluation Pipeline
Configure API keys in EvaluationPipeline/evaluation_config.json first:
- llm.api_base: the API base URL of your model
- llm.model: specify your model name (optional)
- llm.api_key_env: environment variable name for your API key (optional)
cd ../EvaluationPipeline
# Run single question evaluation
python eval_cli.py --mode interactive
# Run batch dataset evaluation
python eval_cli.py --mode batch --dataset sample --workers 20
MedResearcher-R1 Demo Video
xbench_demo
Citations
@article{medresearcher2025,
title={MedResearcher-R1: Expert-Level Medical Deep Researcher via A Knowledge-Informed Trajectory Synthesis Framework},
author={\author{Ailing Yu, Lan Yao, Jingnan Liu, Zhe Chen, Jiajun Yin, Yuan Wang, Xinhao Liao,Zhiling Ye,Ji Li,Yun Yue,Hansong Xiao,Hualei Zhou,Chunxiao Guo,Peng Wei,Jinjie Gu},
journal={arXiv preprint arXiv:https://arxiv.org/pdf/2508.14880},
year={2025}
}
Misc
评论