README.md

metadata

language:
  - en
license: apache-2.0
task_categories:
  - text-generation
tags:
  - instruction-tuning
  - supervised-fine-tuning
size_categories:
  - 1M<n<10M

MBD-ComposedSFT

Dataset Summary

This dataset is a composed supervised fine-tuning (SFT) dataset that combines multiple high-quality instruction-following datasets. It comprises approximately 3,679,149 instruction tuning samples covering diverse domains including mathematics, coding, and general instruction following.

Composition & Sources

The dataset is constructed from the following sources:

• smoltalk: 1,043,917 samples
• OpenMathInstruct 2 1M: 999,117 samples
• tulu 3 sft mixture: 937,730 samples
• opc sft stage2: 436,347 samples
• MathInstruct: 262,038 samples

Data Format

Each record follows this structure:

{
    "query": "instruction or question text",
    "response": "response or answer text"
}

Citation

If you use this dataset, please cite the original source datasets:

• MathInstruct
• OpenMathInstruct-2
• Tulu 3
• SmolTalk
• OPC-SFT

License

This dataset is released under the Apache 2.0 license.

Acknowledgement

This dataset is derived from a thorough cleaning and re-processing of the ReFusion dataset. We acknowledge the creators and contributors of ReFusion for their foundational work, on which this dataset is based.