Dataset · Icey444
Million-Scale Reasoning Image Retrieval Dataset
Icey444/retdata
Over 1.7 million source photo and instruction to target photo pairs for multimodal reasoning retrieval training.
What's inside
A large multimodal dataset comprising over 1.7 million quality-filtered examples curated by a judge model from two subsets (web_search and faiss_legacy). It contains pairs of source images, text instructions, and target images linked by logical reasoning rather than visual similarity. Packaged with embedded images in Qwenvl2d5 v2 format.
What you can do with it
- SFT of multimodal models (such as Qwen-VL) for complex image retrieval
- Training composed image retrieval systems with text instruction comprehension
- Training retrievers that rely on logical reasoning instead of visual similarity
- Tasks
- Image → image
- Size
- 1M–10M rows
- Downloads this month
- 45.7K
- Likes
- 0
- Access
- on request on Hugging Face
How to load it
from datasets import load_dataset
ds = load_dataset("Icey444/retdata")