Rayenlat/cpv-match-corpus

CPV Match corpus

Cleaned, deduplicated EU public procurement contract notices from TED, with the chronological, CPV-disjoint and cross-lingual splits used by the CPV Match project.

What this is, and the caveat that comes first

Labels and evaluation both come from CPV codes, the EU procurement vocabulary. A model trained on this corpus is trained on CPV and evaluated against CPV, so what any metric measures is CPV-proxy retrieval quality, not relevance. Nobody judged whether a retrieved notice is useful to a supplier. Read every number in that light.

Construction

• Source: TED Search API, form-type=competition, publication dates 2024-01-01 to 2026-07-31.
• Languages: German, French, Polish, Spanish, Italian, selected by volume and language family.
• Text: buyer-written title (BT-21) plus description (BT-24), in the notice's own language, whitespace normalised, minimum 300 characters.
• notice-title is excluded. TED composes that field as <country> - <CPV label> - <buyer title>, so it contains the CPV label the project predicts. Including it would leak the label into the input.
• Deduplicated with MinHash over word 5-shingles before splitting, so no near duplicate straddles a split boundary.
• Splits are chronological by publication date. No random splitting.

The test split is frozen: it is read once, for final reporting, and never for threshold tuning, hyperparameter selection or hard negative mining.

Counts

Every count and checksum for this build is in build_report.json, produced by the build run that created these files.

Split	Rows
corpus.parquet	408002
splits/train.parquet	351603
splits/dev.parquet	26677
frozen/test.parquet	29722
splits/cpv_disjoint.parquet	67010
splits/cross_lingual.parquet	29506

Built 2026-08-24 from git commit fc1a95bb9fee0ad2855b73913a0ac45b4670170e.