ESCI-S Dataset
Overview
ESCI-S is a community-built extension of the Amazon ESCI Dataset that adds richer product metadata to the original query–product pairs. While the original ESCI release includes basic product fields, ESCI-S enriches each product record with additional attributes useful for training and evaluation.
What It Adds
- Extended product metadata beyond the original ESCI fields (titles, descriptions, bullets)
- Additional structured attributes that support feature engineering
- Useful for experiments where richer product context improves model quality
Use Cases
- Training embedding and ranking models that benefit from richer product features
- Ablation studies on the effect of metadata richness on retrieval quality
- Augmenting the base Amazon ESCI Dataset for more nuanced experiments
Relationship to ESCI
ESCI-S reuses the same query–product pairs and ESCI relevance labels from the original dataset. It does not change the labels — it only enriches the product-side metadata. Evaluations remain directly comparable to the base ESCI benchmark.
Known Issues — Image Link Rot
The extended metadata includes product image URLs, but a large share of them are no longer usable in practice:
- 131,054 products in the extended dataset have no image at all.
- Of the products that do carry an image URL, roughly 46% of the links are dead — measured by Andrew Kornilov while downloading two independent samples of ~20,000 products: 9,195 and 9,216 failures respectively.
Practical consequence: image URLs in ESCI-S cannot be trusted as-is. Anyone building an image-based pipeline has to download and validate every image before indexing, and over-sample to hit a target corpus size. In How to Evaluate Image Search in Qdrant Using Quepid Part 1, sampling 40,000 products yielded only 21,589 with a working image.
The text-side metadata and the ESCI relevance labels are unaffected — this only bites Multimodal Embeddings / image-search work.
Related Concepts
- Amazon ESCI Dataset — the base dataset this extends
- Judgment Lists — shared annotation structure
- Semantic Search — primary model type evaluated with ESCI-based data
- Learning to Rank — downstream task using these labels
- How to Evaluate Image Search in Qdrant Using Quepid Part 1 — hands-on use of the image metadata, and where the link-rot numbers come from