bpiwowar/abo-matching

ABO subset — product matching (teaching)

Subset of Amazon Berkeley Objects (CC BY 4.0), repackaged for a practical on vision-language product matching: « are these two listings the same product? »

⚠️ Positive pairs are simulated. ABO has no cross-seller duplicates, so a positive pair is two catalog views of the same listing. Real marketplace duplicates are two different sellers photographing the same object — harder, and what the industrial pipeline (embedding for recall, threshold, VLM for the grey band) is actually built for.

Configurations

listings — one row per product, image_main / image_other are two views:

column
item_id	ABO listing id
item_name, bullet_points, brand, color, product_type	metadata (fr_FR, fr_CA, fr_BE, en_US, …)
text	title + brand + colour, ready for text-image matching
image_main, image_other	two views, 224 px max side
image_main_id, image_other_id	ABO image ids

pairs — labelled pairs, referring to listings by (item_id, image_id):

column
left_item_id, left_image_id, right_item_id, right_image_id	the two views compared
left_text	the left listing's title (+ brand, colour)
right_text	the right listing's title as another seller would have written it (brand dropped, word window, colour sometimes appended) — so a positive pair does not share the identical string, and the text score stops being a leak; built the same way whatever the label
label	1 = same product, 0 = different
hardness	same_listing, same_type_same_brand, same_type, random
product_type	of the left listing

The negative mix is drawn to 35% same_type_same_brand, 15% same_type, 50% random — left to itself the sampler would return ~70% same-brand negatives (many ABO listings share an Amazon private brand), which leaves no usable operating point.

listings = load_dataset("bpiwowar/abo-matching", "listings", split="test")
pairs = load_dataset("bpiwowar/abo-matching", "pairs", split="test")
views = {  # (item_id, image_id) -> PIL image
    (r["item_id"], i): im
    for r in listings
    for i, im in ((r["image_main_id"], r["image_main"]), (r["image_other_id"], r["image_other"]))
}

Modifications (as required by CC BY 4.0)

Selection of listings with at least two catalog views, a brand and a product_type; round-robin balancing across product_type (222 types, product_type in ABO is heavily skewed); listings whose title is in none of fr_FR, fr_CA, fr_BE, en_US, … dropped, and one metadata value kept per field following that preference order; images downscaled to 224 px max side and re-encoded as JPEG; pair table built as described above.

Split	Listings	Pairs
train	16094	32188
test	3906	7812

Attribution

Data, including all images: Amazon.com. Dataset, archives and benchmark sets: Matthieu Guillaumin (Amazon.com), Thomas Dideriksen (Amazon.com), Kenan Deng (Amazon.com), Himanshu Arora (Amazon.com), Jasmine Collins (UC Berkeley) and Jitendra Malik (UC Berkeley) — Amazon Berkeley Objects, CC BY 4.0.

Embeddings

L2-normalised google/siglip2-base-patch16-224 embeddings, precomputed so a 45 min session does not spend its time on a forward pass: image-embeddings (one row per view, item_id + image_id + embedding) and text-embeddings (one row per listing, over text). Titles are lowercased and stripped of punctuation first — SigLIP 2 dropped the canonicalisation SigLIP 1 did inside its tokenizer, and raw titles cost ~80 points of top-1 retrieval.