MathX-20M: A High-Quality Corpus for Mathematical Instruction Tuning

Modotte's MathX-20M dataset is a 20-million-instance corpus curated from public sources and enhanced with synthetic data from both closed and open-source models. It is designed as a foundation for instruction-based model tuning and fine-tuning, focusing on mathematical reasoning. The dataset was last updated on February 10, 2026.

Use Cases

• Instruction tuning of language models based on the described high-quality, instruction-based corpus.
• Improving mathematical reasoning capabilities in LLMs based on the dataset's focus on mathematics.
• Generating synthetic training data for math QA tasks based on the described use of synthetic data generation.
• Benchmarking model performance on math-related text generation tasks.

Strengths

• Dataset size is explicitly stated as 20 million instances.
• Data is described as curated from high-quality public sources and enhanced with synthetic data.
• Last update metadata is precise: 2026-02-10 07:23:59.

Limitations

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

Provenance

Source

Modotte, via Hugging Face.

Collection Method

Curated from high-quality public sources and enhanced with synthetic data from closed and open-source models.

Time Range

null

Freshness

Last updated 2026-02-10 07:23:59.

Geography

null