Mini-R1: Synthetic Math Reasoning Data for Post-Training

A synthetic math reasoning dataset generated by DeepSeek-V4-Flash for a lightweight R1-style post-training experiment based on Qwen2.5-1.5B-Instruct. It includes raw teacher-generated reasoning-chain data and SFT training data. The dataset was created by author z2h07 and last updated on 2026-06-12.

Use Cases

• Fine-tuning small language models for math reasoning based on synthetic reasoning-chain data.
• Studying the distillation of reasoning capabilities from larger to smaller models.
• Reproducing R1-style post-training experiments on single-GPU hardware.

Strengths

• Dataset is specifically designed for a lightweight, single-GPU reproduction project.
• Includes both raw teacher-generated data and processed SFT training data.
• Generated using a specific, named model (DeepSeek-V4-Flash) for a specific base model (Qwen2.5-1.5B-Instruct).

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and total size are unknown, which may limit suitability assessment.

Provenance

Source

huggingface

Collection Method

Synthetically generated by DeepSeek-V4-Flash.

Freshness

Last updated 2026-06-12 09:10:28; freshness should be verified.