Magpie Reasoning V2: 250,000 CoT Traces from DeepSeek R1 and Llama 70B

Magpie-Align released this dataset in January 2025, providing 250,000 synthetic reasoning records for large language model alignment. It utilizes the Magpie method to generate Chain-of-Thought (CoT) data from DeepSeek-R1 and Llama-3-70B without requiring manual prompt engineering.

Use Cases

• Fine-tuning models for complex logic using the Chain-of-Thought (CoT) reasoning traces
• Instruction alignment training using the synthetic instruction-response pairs
• Distilling reasoning capabilities from DeepSeek-R1 into smaller open-source models

Strengths

• 250,000 reasoning records
• Chain-of-Thought (CoT) formatting for multi-step logic
• Generated using high-parameter models including Llama-3-70B and DeepSeek-R1

Limitations

• Synthetic data nature may result in unverified logical steps or hallucinations
• Potential bias inherited from the teacher models (DeepSeek-R1 and Llama-3-70B)

Provenance

Source

Magpie-Align (Arxiv 2406.08464)

Collection Method

Synthetic generation via the Magpie method extracting model internal states

Time Range

2024-2025

Freshness

Last updated January 27, 2025.