MegaScience: 1.25 Million Instances for Scientific Reasoning Post-Training

MegaScience contains 1.25 million instances for scientific reasoning post-training, released by GAIR-NLP in July 2025. The collection aggregates multiple public datasets using data selection methods optimized through ablation studies to improve model performance in STEM domains.

Use Cases

• Fine-tuning large language models using the 1.25 million science reasoning instances
• Evaluating model logic on STEM-specific text generation tasks
• Researching data selection strategies for post-training datasets

Strengths

• 1.25 million instances
• Ablation-study-driven selection
• STEM-focused reasoning tasks

Limitations

• Aggregated from public sources which may overlap with model pre-training data
• Restrictive CC BY-NC-SA 4.0 license prohibits commercial use

Provenance

Source

GAIR-NLP (GitHub: GAIR-NLP/MegaScience)

Collection Method

Aggregated from multiple public datasets with optimized selection

Freshness

Last updated July 2025.