APIGen_MT_Pipeline/
-
This directory contains the data synthesis pipeline code based on the APIGen_MT (https://arxiv.org/abs/2504.03601), adapted to the BFCL environment. It enables the generation of synthetic training data for further use in downstream models.
-
The core logic and execution scripts are included.
-
For detailed usage instructions and command-line execution, please refer to the README.md file located inside this directory.
Progra_RAGEN/
-
This directory houses the core implementation of the Progra algorithm, a Progress-Awareness Guided RL method built upon the RAGEN (https://github.com/RAGEN-AI/RAGEN) framework. The algorithm has been adapted and applied to both BFCL and Tau-Bench environments.
-
Note: Due to the large and complex structure of RAGEN, unmodified RAGEN code is omitted here.
-
Users are expected to download the official RAGEN repository separately and install the required environment accordingly.
- Clone RAGEN Repository from https://github.com/RAGEN-AI/RAGEN, using commit 5a59e2c5d6f92ccaae3a21202420f1b0c34d39b3 to ensure compatibility
- Install RAGEN Environment. Follow the installation instructions provided in the RAGEN repository.
- Prepare Datasets. Copy the Progra_Data/ folder to the expected dataset directory inside RAGEN or modify the paths in the configuration files.
- Prepare Env in RAGEN. Copy all sub folders in Progra_RAGEN/ragen/ and Progra_RAGEN/config/ to corresponding path in original RAGEN dir, including exchanging Progra_RAGEN/ragen/llm_agent/agent_proxy.py and other files.
- Run Training Process By:
python ./train.py --config-name CONFIG_NAME
CONFIG_NAME in our example is Qwen2.5-7B-Instruct_BFCL_GRPO_summary_detach_optimization_action_and_summary