Practical overview
When it fits
Use JustGRPO when you need research code for GRPO-based training or benchmark evaluation of diffusion language models, rather than a hosted chat or inference service.
Who it is for
- Researchers studying reinforcement learning for diffusion language models.
- Practitioners who want to train or evaluate dLLMs on GSM8K, MATH-500, HumanEval or MBPP.
Documented uses
- Train a diffusion language model with GRPO using autoregressive order.
- Evaluate project checkpoints on supported math and coding benchmarks.
- Compare full fine-tuning and LoRA results using the provided checkpoints and workflows.
Requirements & limits
The repository is public and specifies an MIT license.
- The documented training examples use eight processes and set global batch size to 64.
- Code training requires the AceCode-Hard dataset file at datasets/acecode_hard.jsonl.
- The README provides both full fine-tuning checkpoints and LoRA adapters, and says evaluation detects LoRA adapters.
- The project uses autoregressive order during RL training while retaining parallel decoding at inference.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Install the JustGRPO repository
Prepare the repository and install its listed Python dependencies.
Before you start
- A Linux or server environment.
- Python and the repository's documented dependencies.
Steps
Clone https://github.com/LeapLabTHU/JustGRPO.git.
Change into the JustGRPO directory.
Install requirements with pip install -r requirements.txt.
Expected outcome
The repository is cloned and its requirements file has been installed.
Guide limitations
- This setup sequence does not cover training or evaluation.
- The README lists Accelerate, Transformers, Datasets and evaluation utilities among the dependencies.
Run a documented benchmark evaluation
Use the repository's evaluation script with a listed task and model checkpoint.
Before you start
- The JustGRPO repository and its installed requirements.
- A compatible project checkpoint and the base model it requires.
- A multi-process environment for the documented eight-process command.
Steps
Choose a listed task: gsm8k, math500, humaneval or mbpp.
Provide the checkpoint location as the --ckpt_path value.
Run eval.py with torchrun --nproc-per-node=8 and the documented generation length, steps and block length arguments.
Expected outcome
The evaluation script runs for the selected task and checkpoint using the configured generation parameters.
Guide limitations
- The README's command is an example; it does not establish performance on a different setup.
- Checkpoint files and the needed base model must be available at the paths used.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 14 October 2026
- Original website title
- LeapLabTHU/JustGRPO · Official homepage title · retained in its source language