Euclid: Supercharging Multimodal LLMs with Synthetic High-Fidelity Visual Descriptions

Dataset Card for Geoperception

A Benchmark for Low-level Geometric Perception

Dataset Details

Dataset Description

Geoperception is a benchmark focused specifically on accessing model's low-level visual perception ability in 2D geometry.

It is sourced from the Geometry-3K corpus, which offers precise logical forms for geometric diagrams, compiled from popular high-school textbooks.

Dataset Sources

• Repository: https://github.com/euclid-multimodal/Euclid
• Paper: https://arxiv.org/abs/2412.08737
• Demo: https://euclid-multimodal.github.io/

Uses

Evaluation of multimodal LLM's ability of low-level visual perception in 2D geometry domain.

Dataset Structure

Fields

• id identification of each data instance
• question question
• answer answer
• predicate question type, including
  • PointLiesOnLine
  • LineComparison
  • PointLiesOnCircle
  • AngleClassification
  • Parallel
  • Perpendicular
  • Equal
• image image

Evaluation Result

Model	POL	POC	ALC	LHC	PEP	PRA	EQL	Overall
Random Baseline	1.35	2.63	59.92	51.36	0.23	0.00	0.02	16.50
Open Source
Molmo-7B-D	11.96	35.73	56.77	16.79	1.06	0.00	0.81	17.59
Llama-3.2-11B	16.22	37.12	59.46	52.08	8.38	22.41	49.86	35.08
Qwen2-VL-7B	21.89	41.60	46.60	63.27	26.41	30.19	54.37	40.62
Cambrian-1-8B	15.14	28.68	58.05	61.48	22.96	30.74	31.04	35.44
Pixtral-12B	24.63	53.21	47.33	51.43	21.96	36.64	58.41	41.95
Closed Source
GPT-4o-mini	9.80	61.19	48.84	69.51	9.80	4.25	44.74	35.45
GPT-4o	16.43	71.49	55.63	74.39	24.80	60.30	44.69	49.68
Claude 3.5 Sonnet	25.44	68.34	42.95	70.73	21.41	63.92	66.34	51.30
Gemini-1.5-Flash	29.30	67.75	49.89	76.69	29.98	63.44	66.28	54.76
Gemini-1.5-Pro	24.42	69.80	57.96	79.05	38.81	76.65	52.15	56.98

Citation

If you find Euclid useful for your research and applications, please cite using this BibTeX:

@article{zhang2024euclid,
  title={Euclid: Supercharging Multimodal LLMs with Synthetic High-Fidelity Visual Descriptions},
  author={Zhang, Jiarui and Liu, Ollie and Yu, Tianyu and Hu, Jinyi and Neiswanger, Willie},
  journal={arXiv preprint arXiv:2412.08737},
  year={2024}
}