MAmmoTH-VL-Instruct-12M: Visual Instruction-Tuning Data for Math and Science

MAmmoTH-VL-Instruct-12M is a dataset of 12 million visual instruction examples created by the MAmmoTH-VL organization. The data was generated through a pipeline involving manual source collection, rewriting by MLLMs/LLMs, and filtering by an MLLM judge, with examples focused on math and science. It was last updated on January 5, 2025.

Use Cases

• Fine-tuning vision-language models for detailed, step-by-step problem-solving based on the described instruction-tuning pipeline
• Training models to generate educational explanations in math and science based on the described data categories
• Benchmarking model performance on visual instruction-following tasks based on the described MLLM-judged filtering process
• Studying the effect of data rewriting and filtering on model reasoning capabilities based on the described pipeline

Strengths

• Contains 12 million instruction examples, providing a large-scale resource for training
• Data underwent a described three-step pipeline involving MLLM/LLM rewriting and filtering for quality control
• Examples are specifically focused on math and science categories, offering domain-specific content

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

MAmmoTH-VL organization

Collection Method

Generated via a pipeline of manual source collection, MLLM/LLM rewriting, and MLLM-based filtering

Freshness

Last updated 2025-01-05 03:53:38; freshness should be verified