https://github.com/TIGER-AI-Lab/OpenResearcher

OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis
https://github.com/TIGER-AI-Lab/OpenResearcher
deep-research llm retrieval
Last synced: 3 months ago
JSON representation
OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis
Host: GitHub
URL: https://github.com/TIGER-AI-Lab/OpenResearcher
Owner: TIGER-AI-Lab
Created: 2026-02-03T00:34:05.000Z (4 months ago)
Default Branch: main
Last Pushed: 2026-02-24T01:20:42.000Z (4 months ago)
Last Synced: 2026-02-24T03:01:58.212Z (4 months ago)
Topics: deep-research, llm, retrieval
Language: Python
Homepage: https://github.com/TIGER-AI-Lab/OpenResearcher
Size: 7.96 MB
Stars: 386
Watchers: 1
Forks: 41
Open Issues: 0
Metadata Files:
- Readme: README.md
Awesome Lists containing this project

awesome - TIGER-AI-Lab/OpenResearcher - OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis (Python)
README

          


  

  






    

  

    

  

  

    

  

  










  🤗 HuggingFace ｜

 Blog ｜ Slack |  WeChat 



## 📣 News 

+ **[2026.2.25]** 🔥 Honored to be among the **top 3 trending datasets** on 🤗 [Hugging Face](https://huggingface.co/datasets) — now **11K+** downloads! 🚀

+ **[2026.2.18]** 🧪 The OpenResearcher training [code](https://github.com/TIGER-AI-Lab/OpenResearcher?tab=readme-ov-file#-optional-train-your-own-openresearcher) is now available. Start training your own OpenResearcher!

+ **[2026.2.14]** 📸 Excited to have our OpenResearcher demo [video](https://x.com/zhuofengli96475/status/2021682952074097086). Dive in and unlock the power of Deep Research today!

+ **[2026.2.12]**  🔥 Excited to see **OpenResearcher** powering deep research trajectory generation in [**NVIDIA’s NeMo Data Designer**](https://nvidia-nemo.github.io/DataDesigner/latest/devnotes/deep-research-trajectories-with-nemo-data-designer-and-mcp-tool-use/)!

+ **[2026.2.10]** 🚀 Our X [post](https://x.com/DongfuJiang/status/2020946549422031040) received **1.2K+ likes**! Feel free to check out the post and join the discussion! 💬

## 💥 Introduction

**OpenResearcher** is a fully open agentic large language model (30B-A3B) designed for **long-horizon deep research** scenarios. It achieves an impressive **54.8%** accuracy on [BrowseComp-Plus](https://huggingface.co/spaces/Tevatron/BrowseComp-Plus), surpassing performance of `GPT-4.1`, `Claude-Opus-4`, `Gemini-2.5-Pro`, `DeepSeek-R1` and `Tongyi-DeepResearch`. We **fully open-source** the training and evaluation recipe—including data, model, training methodology, and evaluation framework for everyone to progress deep research.



  






## 🏆 Deep Research Benchmark Results



  



## ✨ Features

+ 🔑 **Fully Open-Source Recipe** — We fully open-source our 96K high-quality [DeepResearch trajectory dataset](https://huggingface.co/datasets/OpenResearcher/OpenResearcher-Dataset) with 100+ turns generated by GPT-OSS-120B with [native browser tools](https://docs.vllm.ai/projects/recipes/en/latest/OpenAI/GPT-OSS.html#usage:~:text=Limitation%20section%20below.-,Tool%20Use,-%C2%B6), the leading [30B-A3B model](https://huggingface.co/OpenResearcher/OpenResearcher-30B-A3B) trained on it, [distillation recipe](https://boiled-honeycup-4c7.notion.site/OpenResearcher-A-Fully-Open-Pipeline-for-Long-Horizon-Deep-Research-Trajectory-Synthesis-2f7e290627b5800cb3a0cd7e8d6ec0ea?source=copy_link), and a lightweight [DeepResearch evaluation framework](https://github.com/TIGER-AI-Lab/OpenResearcher) to progress deep research.

+ 💰 **Highly Scalable and Low-Cost** — We generate DeepResearch trajectories at massive scale using self-built retriever over a dedicated ~11B-token [corpus](https://huggingface.co/datasets/OpenResearcher/OpenResearcher-Corpus), eliminating the need for external Search APIs. This scalable retriever significantly reduces training costs.

+ 🚀 **Remarkable Performance on Deep Research Benchmarks** — OpenResearcher demonstrates leading performance across a range of deep research benchmarks, including BrowseComp-Plus, BrowseComp, GAIA, xbench-DeepSearch.

## 📋 Table of Contents

- [🛠 Environment Setup](#-environment-setup)

  - [Installation](#installation)

  - [Deep Research Benchmarks Preparation](#deep-research-benchmarks-preparation)

- [🔍 Configuration](#-configuration)

- [🚀 Quick Start](#-quick-start)

- [🔬 Benchmark OpenResearcher](#-benchmark-openresearcher)

  - [Example 1: BrowseComp-Plus with Local Search Engine](#example-1-browsecomp-plus-with-local-search-engine)

  - [Example 2: GAIA with Serper API (No Local Search Needed)](#example-2-gaia-with-serper-api-no-local-search-needed)

  - [Evaluation](#evaluation)

  - [Quick Commands](#quick-commands)

- [🧪 (Optional) Train Your Own OpenResearcher](#-optional-train-your-own-openresearcher)

- [🤝 Core Contributors](#-core-contributors)

- [🎓 Advisors](#-advisors)

- [🙏 Acknowledgements](#-acknowledgements)

- [✨ Contributing](#-contributing)

- [📚 Citation](#-citation)

## 🛠 Environment Setup

We run this repo on the following setup:

+ 8 * A100 80G Nvidia GPUs

+ Linux operating system

Other hardware setups can also work, but remember to modify the corresponding parameters.

### Installation 

```bash

sudo apt update 

sudo apt install -y openjdk-21-jdk

# install uv

curl -LsSf https://astral.sh/uv/install.sh | sh

uv venv --python 3.12

source .venv/bin/activate

# install tevatron for BrowseComp-plus 

git clone https://github.com/texttron/tevatron.git

cd tevatron

uv pip install -e .

cd ..

# install all dependencies automatically

uv pip install -e .

```

### Deep Research Benchmarks Preparation

Run the setup script to automatically download the **[BrowseComp-Plus](https://arxiv.org/abs/2508.06600)** benchmark. Other benchmarks, including **[BrowseComp](https://arxiv.org/abs/2504.12516)**, **[GAIA](https://arxiv.org/abs/2311.12983)** and **[xbench-DeepResearch](https://github.com/THUDM/xbench)**, will be set up automatically when they are first used.

```bash

bash setup.sh

```

**This script will:**

- ✅ Verify Python 3.12 virtual environment and automatically install any missing dependencies

- ✅ Downlaod BrowseComp-Plus dataset from HuggingFace and set up the directory structure

For more info about these deep research benchmarks, see [benchmarks.md](assets/docs/benchmarks.md) 

## 🔍 Configuration

Copy the template and configure your API keys:

```bash

cp .env.template .env

```

Edit `.env`:

```bash

# Serper API (for web search when using browser_backend=serper)

SERPER_API_KEY=your_key        # Get from: https://serper.dev/

# OpenAI API (for evaluation scoring)

OPENAI_API_KEY=your_key        # Get from: https://platform.openai.com/api-keys

```

## 🚀 Quick Start

**Prerequisites:** Install dependencies and configure API keys (see [Environment Setup](#-environment-setup) and [Configuration](#-configuration))

1. **Deploy OpenResearcher-30B-A3B**:

```bash

bash scripts/start_nemotron_servers.sh

```

The complete vLLM server logs can be found in the `logs` directory.

2. **Run your first task** (Before proceeding, check the logs in `logs` directory to ensure the vLLM server is deployed.)

```python

import asyncio

from deploy_agent import run_one, BrowserPool

from utils.openai_generator import OpenAIAsyncGenerator

async def main():

    # Initialize generator and browser

    generator = OpenAIAsyncGenerator(

        base_url="http://localhost:8001/v1",

        model_name="OpenResearcher/OpenResearcher-30B-A3B",

        use_native_tools=True

    )

    browser_pool = BrowserPool(search_url=None, browser_backend="serper")

    # Run deep research

    await run_one(

        question="What is the latest news about OpenAI?",

        qid="quick_start",

        generator=generator,

        browser_pool=browser_pool,

    )

    browser_pool.cleanup("quick_start")

if __name__ == "__main__":

    asyncio.run(main())

```

The deep research agent will automatically search the web, browse webpages, and extract relevant information. You'll see the final answer along with all intermediate reasoning steps.

## 🔬 Benchmark OpenResearcher

We benchmark our OpenResearcher-30B-A3B using below deep research benchmarks: 

| Benchmark | Dataset Key | Size | Language | Search Backend | Description |

|-----------|-------------|------|----------|----------------|-------------|

| [BrowseComp-Plus](https://arxiv.org/abs/2508.06600) | `browsecomp_plus` | 830 | EN | local | Deep-research benchmark from BrowseComp isolating retriever and LLM agent effects |

| [BrowseComp](https://arxiv.org/abs/2504.12516) | `browsecomp` | 1,266 | EN | serper | A Simple Yet Challenging Benchmark for Browsing Agents |

| [GAIA-text](https://arxiv.org/abs/2311.12983) | `gaia` | 103 | EN | serper | Text-only subset of GAIA benchmark (dev split) |

| [xbench-DeepResearch](https://github.com/THUDM/xbench) | `xbench` | 100 | ZH | serper | DeepSearch benchmark with encrypted test cases |

For more info about these deep research benchmarks, see [benchmarks.md](assets/docs/benchmarks.md) 

### Example 1: BrowseComp-Plus with Local Search Engine

Complete evaluation using local dense search with browsecomp-plus [corpus](https://huggingface.co/datasets/Tevatron/browsecomp-plus-corpus) and [embeddings](https://huggingface.co/datasets/Tevatron/browsecomp-plus-indexes/tree/main/qwen3-embedding-8b) (**note: only applicable for BrowseComp-Plus**):

```bash

# Terminal 1: Start local Dense search service on port 8000

# Embedding model (Qwen3-Embedding-8B) will be deployed on GPUs 7

bash scripts/start_search_service.sh dense 8000

# Terminal 2: Start vLLM servers (requires 4 GPUs)

# TP=2, deploy 2 servers starting from port 8001 on GPUs 0,1,2,3

bash scripts/start_nemotron_servers.sh 2 8001 0,1,2,3

# Terminal 3: Run agent

bash run_agent.sh results/browsecomp_plus/OpenResearcher_dense 8001 2 browsecomp_plus local OpenResearcher/OpenResearcher-30B-A3B

```

What this does:

- Deploys Dense retriever service on port 8000 as search engine

- Launches 2 vLLM servers (ports 8001, 8002) with TP=2 across 4 GPUs

- Runs deepresearch agent with load balancing across both servers

### Example 2: GAIA with Serper API (No Local Search Needed)

Run with Serper Google Search API (**note: applicable to all benchmarks except BrowseComp-Plus**):

```bash

# Terminal 1: Start vLLM servers (requires 4 GPUs)

bash scripts/start_nemotron_servers.sh 2 8001 0,1,2,3

# Terminal 2: Run agent with serper search backend

bash run_agent.sh results/gaia/OpenResearcher_serper 8001 2 gaia serper OpenResearcher/OpenResearcher-30B-A3B

```

**Browser Backend Options:**

- `local` - Use local BM25/Dense search service (for BrowseComp-Plus)

- `serper` - Use Serper Google Search API (for all other benchmarks)

For other parameters, refer to [parameter.md](assets/docs/parameter.md).

### Evaluation

After running experiments, evaluate results:

```bash

# eval on browsecomp_plus

python eval.py --input_dir results/browsecomp_plus_dense/OpenResearcher_dense

# eval on gaia

python eval.py --input_dir results/gaia/OpenResearcher_serper

```

### Quick Commands

| Scenario | Command |

|----------|---------|

| BrowseComp-Plus (BM25) | `bash scripts/start_search_service.sh bm25 8000` then `bash scripts/start_nemotron_servers.sh 2 8001 0,1,2,3` then `bash run_agent.sh results/browsecomp-plus/OpenResearcher_bm25 8001 2 browsecomp_plus local OpenResearcher/OpenResearcher-30B-A3B` |

| BrowseComp-Plus (Qwen3-8B Dense Embeddings) | `bash scripts/start_search_service.sh dense 8000` then `bash scripts/start_nemotron_servers.sh 2 8001 0,1,2,3` then `bash run_agent.sh results/browsecomp-plus/OpenResearcher_dense 8001 2 browsecomp-plus local OpenResearcher/OpenResearcher-30B-A3B` |

| BrowseComp | `bash scripts/start_nemotron_servers.sh 2 8001 0,1,2,3` then `bash run_agent.sh results/browsecomp 8001 2 browsecomp serper OpenResearcher/OpenResearcher-30B-A3B` |

| GAIA | `bash scripts/start_nemotron_servers.sh 2 8001 0,1,2,3` then `bash run_agent.sh results/gaia 8001 2 gaia serper OpenResearcher/OpenResearcher-30B-A3B` |

| xbench-DeepResearch | `bash scripts/start_nemotron_servers.sh 2 8001 0,1,2,3` then `bash run_agent.sh results/xbench 8001 2 xbench serper OpenResearcher/OpenResearcher-30B-A3B` |

For script parameter explanation, refer to [parameter.md](assets/docs/parameter.md).

**Note:** Don't forget to evaluate your results using:  

```bash

python eval.py --input_dir [INPUT_DIR]

```

## 🧪 (Optional) Train Your Own OpenResearcher

Our [OpenResearcher-30B-A3B](https://huggingface.co/OpenResearcher/OpenResearcher-30B-A3B) is trained using [Megatron-LM](https://github.com/NVIDIA/Megatron-LM) on [openresearcher-dataset](https://huggingface.co/datasets/OpenResearcher/OpenResearcher-Dataset). To get started, clone the `openresearcher` branch of the Megatron-LM repository:

```

git clone -b openresearcher https://github.com/jdf-prog/Megatron-LM.git

```

Then, follow the training instructions [here](https://github.com/jdf-prog/Megatron-LM/tree/openresearcher/examples/openresearcher) to train your own OpenResearcher!

## 🤝 Core Contributors

    

        

            

            


            _{Zhuofeng Li}

        

    

        

        

            

            


            _{Dongfu Jiang}

        

    

    

        

        

            

            


            _{Xueguang Ma}

        

    

    

        

            

            


            _{Haoxiang Zhang}

        

    

    

        

            

            


            _{Ping Nie}

        

    

## 🎓 Advisors

      

        

            

            


            _{Wenhu Chen}

        

    

    

        

            

            


            _{Yu Zhang}

        

    

## 🙏 Acknowledgements



  



## ✨ Contributing

We are truly looking forward to open-source contributions to OpenResearcher! If you’re interested in contributing, collaborating, or reporting issues, please feel free to open an issue or submit a pull request (PR). You can also reach us at [zhuofengli12345@gmail.com](mailto:zhuofengli12345@gmail.com).

We are also looking forward to your feedback and suggestions!

##  📚 Citation

```bibtex

@misc{li2025openresearcher,

  title={OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis},

  author={Zhuofeng Li and Dongfu Jiang and Xueguang Ma and Haoxiang Zhang and Ping Nie and Yuyu Zhang and Kai Zou and Jianwen Xie and Yu Zhang and Wenhu Chen},

  year={2025},

  howpublished={\url{https://www.notion.so/OpenResearcher-A-Fully-Open-Pipeline-for-Long-Horizon-Deep-Research-Trajectory-Synthesis-2f7e290627b5800cb3a0cd7e8d6ec0ea}},

  note={Notion Blog}

}

```

## ⭐ Star History

[![Star History Chart](https://api.star-history.com/svg?repos=TIGER-AI-Lab/OpenResearcher&type=date&legend=top-left)](https://www.star-history.com/#TIGER-AI-Lab/OpenResearcher&type=date&legend=top-left)



  

    ↑ Back to Top ↑
ecosyste.ms

Data

Tools

Indexes

Applications

Experiments

Awesome

https://github.com/TIGER-AI-Lab/OpenResearcher

Awesome Lists containing this project

README
