OpenRFT is an open-source project whose goal is to adapt general Reasoning Foundation Models to domain-specific tasks through Reinforcement Fine-Tuning (RFT).
Purpose OpenRFT is dedicated to addressing the problem of adapting general reasoning models to domain-specific tasks. By leveraging domain-specific samples, it tackles two typical challenges: first, the lack of reasoning step data in domain tasks, and second, the limited number of training samples. With these domain samples, OpenRFT can achieve efficient fine-tuning for domain-specific tasks.
Main Capabilities
- Adapt general reasoning foundation models to domain-specific tasks.
- Perform efficient fine-tuning in the absence of reasoning step data and with limited training samples.
- Provide training and evaluation code.
- Release a technical report (arxiv).
Usage The project's training code relies on the two frameworks OpenRLHF and trl. The repository has updated OpenRFT's training and evaluation code. Specific installation commands, running steps, and configuration methods are not provided in the materials.
Applicable Scenarios Suitable for scenarios that require adapting general reasoning foundation models to domain-specific tasks and that face missing reasoning step data or insufficient training samples.
Limitations Specific information such as supported models, data formats, hardware requirements, performance results, and platform compatibility is not provided in the materials.
License This project is released under the MIT License; see the LICENSE file for details. Using this code or related materials indicates agreement to comply with the terms listed in the license.
