Data Quality Improvement Program
Open-Reasoner-Zero offers a complete solution to data problems:
- 57k high-quality dataset: The preprocessed dataset that comes with the project has been screened through multiple stages and contains:
- 20k GPQA Diamond Standards data
- 15k logical reasoning data
- 22k multi-step decision data
- Customized data processing flow: Available in the src/data_processing directory:
clean_raw_data.py- Raw data cleansinggenerate_synthetic.py- Synthetic data generationquality_filter.py- Quality filtering (PPL threshold set to 2.5 by default)
Extended data program
To add field-specific data:
- build up
custom_data/Catalog to store new data - modifications
config.yamlThe data_mix_ratio parameter controls the data mixing ratio in the - Recommended Interactive Validation of Data Quality with Jupyter Notebook
This answer comes from the articleOpen-Reasoner-Zero: Open Source Large-Scale Reasoning Reinforcement Learning Training PlatformThe
May not be reproduced without permission:AI productivity tools " How to solve the problem of insufficient quality of training data for reinforcement learning?
