FishMath SFT Data: 23,257 LLM-Generated Math Solution Traces

FishMath SFT Data is a synthetic dataset for Supervised Fine-Tuning in mathematical reasoning. It contains 23,257 correct solution traces generated by multiple frontier open-source LLMs, covering competition-level math problems from diverse sources. The dataset was created by SakanaAI and was last updated on May 8, 2026.

Use Cases

• Supervised Fine-Tuning of language models for mathematical reasoning based on the described solution traces.
• Benchmarking model performance on competition-level math problems as referenced in the description.
• Studying the reasoning patterns of frontier open-source LLMs on complex mathematical tasks.
• Training models for the Kaggle AI Mathematical Olympiad 3 - Progress Prize 3 (AIMO 3) project.

Strengths

• Contains 23,257 solution traces, providing a substantial volume of training data.
• Traces are described as 'correct', indicating a focus on high-quality, verified outputs.
• Problems are sourced from 'diverse' competition-level math sources, suggesting variety in problem types.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but the specific file formats and data structure are unknown.
• The dataset's license is unspecified, which may restrict usage.

Provenance

Source

SakanaAI

Collection Method

Synthetic generation by multiple frontier open-source LLMs.

Freshness

Last updated 2026-05-08 06:58:00; freshness should be verified.