Subtext-Decor-90 is a benchmark for evaluating multimodal product understanding and retrieval in home decor and interior design. It targets the queries that traditional keyword and vector retrieval struggle with: intent-heavy phrasing, rich aesthetic descriptors, negation, cultural references, and emotional framing (e.g. "a chair that's comfortable for crying in", "rug that hides cat puke but isn't beige").
The benchmark consists of 90 text queries evaluated… See the full description on the dataset page:
https://huggingface.co/datasets/ontoncom/Subtext-Decor-90.
