Instruction-Tuning Data for LLM Reasoning via Chain-of-Thought

Magpie-Align's 150,000-instruction dataset is designed for aligning large language models, specifically for improving reasoning capabilities. It contains synthetic chain-of-thought data generated using models like DeepSeek-R1 and Llama-70B. The dataset was published in a June 2024 technical report and last updated on the Hugging Face platform in January 2025.

Use Cases

• Supervised fine-tuning of LLMs using the 150,000 instruction-response pairs to improve general instruction-following.
• Training models for chain-of-thought reasoning by leveraging the step-by-step reasoning traces included in the data.
• Benchmarking alignment techniques by comparing performance on this open dataset against models trained on private data like Llama-3-Instruct's.
• Studying the effects of synthetic data generation methods, as the dataset is created using advanced models like DeepSeek-R1.

Strengths

• Contains 150,000 high-quality instruction-response pairs for model alignment.
• Data is generated using advanced models (DeepSeek-R1, Llama-70B) to create complex reasoning traces.
• Technical report (arXiv:2406.08464) provides detailed methodology and validation.

Limitations

• Data is synthetic, generated by models, which may inherit or amplify biases from the source models.
• Specific column structure, data diversity, and potential class balance are unknown from the provided description.
• The scope is predefined by the generation methodology, which may limit task variety.

Provenance

Source

Magpie-Align research project.

Collection Method

Synthetically generated using models like DeepSeek-R1 and Llama-70B, as described in the associated technical report.

Time Range

Creation period centered around 2024, per the technical report.

Freshness

Last updated 2025-01-27; based on a June 2024 technical report.