VisionArena Chat: 200,000 Real-World Image Conversations with Pairwise Preferences

VisionArena-Battle contains 200,000 single and multi-turn chats between users and 45 different Vision Language Models (VLMs), collected via the Chatbot Arena platform. The dataset includes approximately 43,000 unique images and conversations across 138 languages, with question categories such as captioning, OCR, and creative writing. It was created by lmarena-ai and last updated on February 4, 2025.

Use Cases

• Training or fine-tuning VLMs based on real-world user interactions and image conversations.
• Benchmarking model performance and preference ranking based on pairwise votes collected in the arena.
• Analyzing VLM capabilities across diverse question categories like captioning, OCR, and creative writing.
• Studying multilingual and multimodal conversational patterns across 138 languages.

Strengths

• Contains 200,000 conversations, providing a substantial volume of interaction data.
• Includes pairwise preference votes, which are valuable for training and evaluating model alignment.
• Covers 45 different VLMs and 138 languages, offering broad model and linguistic diversity.
• Features approximately 43,000 unique images, ensuring a variety of visual inputs.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset description warns that images may contain inappropriate content.

Provenance

Source

lmarena-ai via Hugging Face.

Collection Method

Collected from user interactions with Vision Language Models on the Chatbot Arena platform.

Time Range

null

Freshness

Last updated 2025-02-04 00:57:01; freshness should be verified.

Geography

null