Fast Math R1 SFT: First-Stage Supervised Fine-Tuning Data for Mathematical LLMs

RabotniKuma's dataset is the first-stage Supervised Fine-Tuning (SFT) data presented in the paper 'A Practical Two-Stage Recipe for Mathematical LLMs: Maximizing Accuracy with SFT and Efficiency with Reinforcement Learning'. It was created for the intensive SFT phase to push a model's mathematical accuracy and was last updated on July 15, 2025. The dataset is associated with a GitHub repository detailing a Kaggle competition solution.

Use Cases

• Conducting supervised fine-tuning of language models for mathematical accuracy based on the described two-stage recipe.
• Benchmarking model performance on mathematical problems before reinforcement learning phases.
• Studying the effect of intensive SFT data on subsequent model efficiency in a reinforcement learning pipeline.

Strengths

• Designed specifically for a documented two-stage training recipe for mathematical LLMs.
• Associated with a published paper and a public GitHub repository providing context.
• Last updated on 2025-07-15, indicating recent maintenance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and data size are unknown, which may limit suitability assessment.
• The license is unspecified, which could restrict usage.

Provenance

Source

RabotniKuma on Hugging Face, associated with a Kaggle competition solution GitHub repository.

Collection Method

Constructed for the Supervised Fine-Tuning phase as described in an academic paper.

Time Range

null

Freshness

Last updated 2025-07-15 05:18:50.

Geography

null