Practical overview
When it fits
Use DARE when an experiment focuses on diffusion language models and benefits from a framework that brings supervised fine-tuning, reinforcement-learning training and evaluation together.
Who it is for
- Researchers developing alignment and reinforcement-learning methods for diffusion language models.
- Developers who need to fine-tune or evaluate supported diffusion language models.
Documented uses
- Fine-tune diffusion language models with supervised examples, including documented parameter-efficient fine-tuning recipes.
- Develop reinforcement-learning training runs and extend supported algorithms.
- Run benchmark evaluation workflows and extend evaluations for diffusion language models.
Requirements & limits
The public GitHub repository displays an Apache-2.0 license.
- Training is built on verl; the evaluation framework is built on OpenCompass.
- The README recommends separate Python 3.10 environments for training and evaluation because of dependency conflicts.
- The documented setup includes FlashAttention for training; selected model workflows also call for SGLang.
- Model setup can require replacing model source files with DARE-provided versions, and datasets must be organized in the expected preprocessed format.
- The project describes itself as work in progress, with model and algorithm support subject to change.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up separate training and evaluation environments
Prepare the documented Python environments and install the framework components for training and evaluation.
Before you start
- Conda and Python 3.10
- A clone of the DARE repository
- For training, the repository requirements and FlashAttention 2.8.3
- For evaluation, the repository's OpenCompass component
Steps
Clone the DARE repository.
Create and activate a Python 3.10 Conda environment for training; install the repository requirements and FlashAttention 2.8.3.
Create a separate Python 3.10 Conda environment for evaluation and install the OpenCompass component in editable mode.
Install additional evaluation dependencies only for the chosen tasks, such as HumanEval or Math.
Expected outcome
Separate environments prepared for DARE training and its OpenCompass-based evaluation workflow.
Guide limitations
- The repository warns of dependency conflicts and recommends separate environments.
- This setup recipe does not cover running a training job or evaluation.
- Task-specific benchmarks require their listed additional dependencies.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 14 October 2026
- Original website title
- yjyddq/DARE · Official homepage title · retained in its source language