Open Deep Research Training

What it does

This project provides a comprehensive training pipeline for building deep research agents. It leverages the ART library to specialize Qwen 2.5 14B for LangChain's open deep research framework, using Group Relative Policy Optimization (GRPO) and Supervised Fine-Tuning (SFT) to reach performance levels that exceed state-of-the-art models like Sonnet 4.

Tools

• collect_sft: Collects high-quality samples for the initial supervised fine-tuning run.
• run_sft: Executes the SFT training to establish a strong baseline performance.
• run_train: Runs the main RL training process using GRPO to optimize research capabilities.
• benchmark_model: Evaluates the trained agent using the DeepResearch Bench framework.

Installation

To use this as a server environment, ensure you have uv installed and run:

uv sync

Configuration for Claude Desktop would typically involve pointing to the executed python environment:

{
  "mcpServers": {
    "open-deep-research": {
      "command": "uv",
      "args": ["run", "python", "-m", "open_deep_research_training"]
    }
  }
}

Supported hosts

• claude