OpenThoughts-Agent: Data Recipes for Agentic Models

OpenThoughts-Agent is a large-scale research project that provides optimal tools and data recipes for training small intelligent agent models. It supports large-scale data generation (Datagen), Supervised Fine-Tuning (SFT), and Reinforcement Learning (RL) tasks, seamlessly integrating with high-performance computing clusters through a highly modular HPC launcher. OpenThoughts-Agent offers customizable data generation scripts and training configurations, deeply integrating leading machine learning tools such as Llama Factory, vLLM, and Harbor. It places special emphasis on deploying containerized tools via Harbor and supports remote container providers to adapt to various HPC environments.

Installing OpenThoughts-Agent

Environment Setup

First, create a clean Python 3.12 virtual environment using a common environment management tool (recommended: conda + mamba).
From the project root directory, execute the following command to install core HPC (High‑Performance Computing) and data infrastructure dependencies:

pip install .

The project has many dependencies; it is recommended to use a package manager like uv within the virtual environment.

Optional Extensions Installation

HPC Data Generation Runtime (supports Ray clusters + vLLM services):

pip install .[datagen]

SweSmith Custom Data Generation Tools (adds custom tools on top of the basic datagen extension):

pip install .[datagen,datagen-swesmith]
# If the basic datagen extension is already installed, you can run directly:
# pip install .[datagen-swesmith]

SFT-Related Installation

Initialize git submodules:

git submodule update --init --recursive sft/llamafactory

Install LLaMA Factory directly from the submodule (OT‑Agent does not provide an [sft] extension):

cd sft/llamafactory
pip install -e .[train,liger-kernel,deepspeed] # Select extension components as needed
cd -

The training configuration files compatible with OT‑Agent are located under the sft/lf_configs/** path. For detailed parameters and dependency instructions, please refer to sft/llamafactory/README.md.