Skip to content

mwp_ReFT

Fine-tune language models with reinforcement learning

Visit official site

What mwp_ReFT is

The project is described as an open-source research effort that uses deep reinforcement learning to fine-tune large language models for better task performance, with code and data for reproducing paper results.

What you can do with it

  • Run SFT and ReFT fine-tuning
  • Use Online-SL and Offline-SL self-learning
  • Evaluate with Top-1, Voting@100 and Reranking

Getting started

  • Clone the repository from GitHub
  • Install dependencies following the README file
  • Choose a pre-trained model and checkpoints
  • Run the matching shell script for training or evaluation

Application scenarios

  • A researcher fine-tunes Codellama on GSM8k
  • A developer fine-tunes Galactica on mathqa
  • A team runs ReFT and ReFT-Rerank on svamp

Pros & cons

✓ What we like

  • Pre-trained model checkpoints let users start from several stages
  • Code and data are provided to reproduce results

! What to watch out for

  • Training follows shell scripts and README instructions
  • The repository must be cloned before use

FAQ

Which models is it based on?

Codellama and Galactica are named, under the Apache 2.0 licence.

What evaluation is supported?

Top-1, Voting@100 and Reranking are listed.

Last reviewed: 2026-11-15

More AI coding tools tools

View all →

How we review