- Improve dataset card: Add task categories, tags, and link to survey paper (ed9a1ba0605d604d1f2f0d270963b9d0fcbed164)
Co-authored-by: Niels Rogge <nielsr@users.noreply.huggingface.co>
# GeoGeo
Paper: [Enhancing the Geometric Problem-Solving Ability of Multimodal LLMs via Symbolic-Neural Integration](https://arxiv.org/pdf/2504.12773)
GitHub: [ycpNotFound/GeoGen: A pipeline for the automatic construction of geometry problems along with step-by-step solutions.](https://github.com/ycpNotFound/GeoGen)
- GeoExpand includes 45,526 Q&A samples, generated from 4849 images in total of Geometry3K and PGPS9K.
- GeoSynth includes 62,868 Q&A samples, which one diagram for one Q&A each.
# GeoGen
Dataset of GeoExpand and GeoSynth for paper: [Enhancing the Geometric Problem-Solving Ability of Multimodal LLMsvia Symbolic-Neural Integration](https://arxiv.org/abs/2504.12773)
GitHub: [GeoGen](https://github.com/ycpNotFound/GeoGen)
## Features
- Build high-quality training data for internal MLLM to solve the problem of the lack of natural language CoT (Chain-of-Thought) explanation of existing data.
- GeoExpand focuses on expanding two classic geometry datasets: Geometry3K and PGPS9K.
- GeoSynth contains Q&A pairs and geometry diagrams that synthetized from scratch by random sampling predicates (e.g. geometric entity, `Triangle` or `Square`, and geometric relationship, `IsMidpointOfLine`).
## Statistics
From 4849 images in Geometry3K and PGPS9K, we generate 45,526 QA pairs, forming the GeoExpand dataset.
Using the Plotter module in GeoGen, we synthesize new geometric diagrams and generate CoT data in the same manner to construct GeoSynth:
- Sampled from 3,159 uniqu…
