NodeNovaOrg/bidlens-open-data

BidLens Open Data

How do public buyers in the UK and EU score bids, and who wins them? This dataset harmonises six official feeds from three public procurement services into ten linked Parquet tables. It covers EU-wide TED notices from 2005 and the UK's Find a Tender and Contracts Finder services. Each published award criterion carries its weight, normalised to a share of its lot. Criteria join to the award outcome of the same lot (winner, number of bids, price range) and to resolved buyer and supplier identities. UK suppliers carry Companies House numbers.

Built by NodeNova as the open layer of BidLens. Snapshot v2026.10, CC BY 4.0.

• Hugging Face: https://huggingface.co/datasets/NodeNovaOrg/bidlens-open-data
• GitHub (card, schema, examples): https://github.com/nodenova/bidlens-open-data

Why another procurement dataset

Raw procurement feeds are already open, and several good aggregations exist (the OCP Data Registry, OpenTender, the GTI global dataset). None of them links these three things:

1. Weighted award criteria per lot. A 60/40, a 600/400 and a 60-points scheme all become one weight_pct scale, and criterion_roles corrects the many UK TED notices that list "Price" as a quality criterion.
2. The outcome of that lot: winner, n_bids, lowest and highest bid, award value against estimate.
3. Who the buyer and winner really are. buyers and suppliers resolve thousands of spellings to one identity. UK winners carry the Companies House number, so you can join incorporation date and status from the register.

Quick start

import duckdb

con = duckdb.connect()
# Mean weight of price criteria, per UK authority (UK-native feeds only, see caveats)
con.sql("""
    SELECT b.canonical_name,
           round(avg(c.weight_pct) FILTER (WHERE r.role = 'price'), 1) AS mean_price_weight_pct,
           count(DISTINCT c.notice_id) AS notices
    FROM 'hf://datasets/NodeNovaOrg/bidlens-open-data/data/criteria/*.parquet' c
    JOIN 'hf://datasets/NodeNovaOrg/bidlens-open-data/data/buyers/*.parquet' b
      ON b.buyer_name = c.buyer_name AND b.buyer_id IS NOT NULL
    LEFT JOIN 'hf://datasets/NodeNovaOrg/bidlens-open-data/data/criterion_roles/*.parquet' r
      ON r.name = c.name
    WHERE c.source IN ('fts_bulk_xml', 'fts_api') AND c.weight_pct IS NOT NULL
    GROUP BY b.buyer_id, b.canonical_name HAVING notices >= 10 ORDER BY notices DESC LIMIT 20
""").show()

from datasets import load_dataset

uk_criteria = load_dataset(
    "NodeNovaOrg/bidlens-open-data", "uk_criteria", split="train", streaming=True
)

The default view is UK: uk_criteria, uk_notices and uk_outcomes hold the rows from the three UK-native feeds plus the GB rows of TED. criteria, notices and outcomes hold everything, UK and EU. On disk the split is in the file names (criteria-uk-*.parquet, criteria-other-*.parquet), so a glob picks either half or both.

More in examples/. For serious work, download once (hf download NodeNovaOrg/bidlens-open-data --repo-type dataset --local-dir bidlens-open-data) and query the local files.

Tables

Table	Grain	Rows	of which UK	Size
notices	one notice; one notice per lot on ted_bulk	25,452,313	1,800,414	740 MB
criteria	one published award criterion (per lot)	34,828,647	2,484,224	580 MB
criterion_roles	one raw criterion name → what it is (price, quality, social_value, …)	157,014	UK only	2 MB
outcomes	one award result per winner; a lot with bid statistics but no award has a NULL winner	34,374,398	1,629,001	354 MB
buyers	one raw buyer_name spelling → resolved buyer_id	48,062	UK only	2 MB
suppliers	one raw winner_name spelling → resolved supplier_id and company number	277,993	UK only	8 MB
frameworks	one multi-supplier award notice, ceiling read once	37,506	UK only	2 MB
buyer_routes	one (authority, CPV division): competed / direct / framework award counts	53,727	UK only	1 MB
documents	one tender-document link cited by a notice	491,621	UK only	19 MB
tender_documents	one document actually opened, with its extracted marking-scheme facts	7,767	UK only	1 MB

Snapshot v2026.10, exported 2026-10-02, 1.71 GB of zstd Parquet.

Column-by-column documentation is in SCHEMA.md. manifest.json lists every file with its row count and SHA-256.

Coverage by source

source	Notice rows	Criteria	Outcomes	First published	Last published
contracts_finder	633,109	—	575,305	2016-11-18	2026-09-30
fts_api	5,706	13,557	7,132	2021-01-01	2021-07-27
fts_bulk_xml	319,799	716,258	483,918	2021-01-01	2026-09-07
ted_api	2,045,235	10,271,001	21,362,084	2023-11-02	2026-09-08
ted_bulk	22,419,541	23,707,142	11,819,941	2005-08-06	2023-12-27
ted_xml	28,923	120,689	126,018	2024-01-02	2024-06-28

What each source is:

source	Feed	Licence
ted_bulk	TED bulk CSV exports (contract and award notices, pre-eForms)	EU open data, Decision 2011/833/EU
ted_api	TED eForms search API	EU open data, Decision 2011/833/EU
ted_xml	TED monthly XML packages: notices still published on the pre-eForms forms, Jan–Jun 2024. The API carries these as headers only, so they replace the ted_api row	EU open data, Decision 2011/833/EU
fts_bulk_xml	Find a Tender daily bulk (OCDS and legacy XML)	OGL v3.0
fts_api	Find a Tender OCDS API (2021 only; mostly a duplicate of fts_bulk_xml)	OGL v3.0
contracts_finder	Contracts Finder OCDS. Notices, outcomes and document links; no award criteria	OGL v3.0

TED reaches the dataset by three routes, one per era. ted_bulk runs to the end of 2023; ted_api takes over in Nov 2023, when eForms became mandatory; ted_xml covers the notices still published on the old forms in the first half of 2024. Each TED notice appears once: where two routes carried the same notice, the copy with its criteria and outcomes was kept.

Read this before analysing

Each of these produces a plausible wrong number with no error. All of them were found the hard way while building the corpus.

• Jurisdiction is a property of source, not country. country is the buyer's address and is NULL on every Contracts Finder and fts_api row. For a UK slice, use source IN ('fts_bulk_xml','fts_api','contracts_finder') OR country = 'GB'.
• criterion_type records which TED column a criterion came from, not what it is. For the real role, join criterion_roles on the raw name and fall back to criterion_type where the role is unknown: CASE WHEN r.role IS NOT NULL AND r.role <> 'unknown' THEN r.role WHEN c.criterion_type IN ('price','quality') THEN c.criterion_type ELSE 'unknown' END. For UK weighting benchmarks, use the UK-native feeds. On ted_bulk, roughly half of UK groups list price only inside the quality column.
• Not every criterion is scored. The roles template_slot, not_competed, undisclosed and unsplit are not marks. Leave them out of any price/quality split. Drop the whole lot when not_competed carries weight, because nobody competed for it.
• weight_pct is normalised per lot group, so it always sums to ~100. Test whether a scheme is complete on weight_value instead. On ted_api, weight_pct is NULL where the feed carries no lot identity; that NULL does not mean no weights were published.
• Joining criteria to outcomes: ON o.notice_id = c.notice_id AND o.lot_id = c.lot_id, with both lot_id IS NOT NULL. Never IFNULL(lot_id,''): it joins every criterion to every award in the notice.
• Money has no currency on most of the EU estate. currency and value_currency are empty on every ted_api and ted_xml row and on most fts_bulk_xml legacy rows, and those values are in the notice's own currency: euro, złoty, lei, forint, all in one column. Only ted_bulk (converted to EUR) and contracts_finder (GBP) are safe to sum or average as they stand. Elsewhere, compare within one country, or filter currency IS NOT NULL.
• Never SUM(award_value) over multi-winner notices. A framework publishes one ceiling and repeats it on every winner's row. Use frameworks (ceiling, read once), filter shape = 'framework', and divide by n_suppliers, not n_award_rows.
• ted_bulk rows are notice × lot. Count DISTINCT notice_id when you mean notices. ted_api and ted_xml hold one row per notice, so a COUNT(*) of TED notices by year falls about threefold at the 2023/2024 switch while the real volume does not.
• A TED series across 2023–2024 needs all three TED sources. Filter source IN ('ted_bulk','ted_api','ted_xml'). Leaving out ted_xml drops about half of January 2024's award notices, which were published on the old forms.
• Group on buyer_id / supplier_id, never on the raw names. Join the lookup tables on the exact raw string (b.buyer_name = c.buyer_name), without case-folding, and filter … IS NOT NULL. A --- in buyer_name is a list of co-buyers. method is the resolution tier: on the supplier side, registry means a real company number. suppliers covers the UK-native feeds only.
• The same UK procurement can appear twice. Above-threshold contracts go to both Find a Tender and Contracts Finder under different ocid namespaces, roughly 12% of FTS award notices.
• Text columns are not corpus-wide. title covers about a third of ted_bulk and none of it before 2009. notices.description is empty on ted_api, but criteria.description is 98.7% filled there. Treat any text-search count as a floor.
• published has two shapes (2024-05-20… and 20240520). Use try_strptime(replace(left(published,10),'-',''),'%Y%m%d').
• Open tenders need stage = 'tender', not a date filter. Planning notices carry dates that look like deadlines. deadline is populated on the OCDS feeds and about half of ted_api, and is NULL on ted_bulk and ted_xml.
• documents are links; tender_documents are documents that were actually read. One is not the other's denominator.

What was changed or left out, and why

• Email addresses are replaced with [email] in every text column except url. The feeds embed officials' contact addresses in free text, and bulk republication is a separate act under UK GDPR. Per-column counts are in manifest.json (emails_redacted). The url column links back to the original notice. Addresses written out to dodge spam filters (jane.smith at council dot gov dot uk) are not recognised and can remain: a few hundred, in notice descriptions.
• Tender document text is not included. tender_documents keeps only the facts extracted from each pack (band count, scales, band keywords, rubric score). The OGL covers what an authority authored, and an ITT pack often contains third-party material.
• Not included: FOI-released evaluation material (no redistribution licence, and it contains personal data), and the Companies House register itself (join it from source on suppliers.company_number).
• Sole traders can appear as winner_name, exactly as the publishing authority released them. To ask for a correction or removal of your own details, email info@nodenova.co.uk; for anything else, open an issue. Accepted requests are applied at the next snapshot.

Licence and attribution

This dataset is released under CC BY 4.0. The code in examples/ and tools/ is MIT (LICENSE). If you use it, please credit "BidLens Open Data, NodeNova", and keep these source attributions:

• Contains public sector information licensed under the Open Government Licence v3.0 (Find a Tender, Contracts Finder).
• Source: Tenders Electronic Daily, © European Union, https://ted.europa.eu, reused under Commission Decision 2011/833/EU.

NodeNova is not affiliated with or endorsed by the Cabinet Office, the Government Commercial Agency or the Publications Office of the EU.

Citation

@dataset{nodenova_bidlens_open_data_2026,
  title     = {BidLens Open Data: UK and EU public procurement award criteria and outcomes},
  author    = {{NodeNova}},
  year      = {2026},
  version   = {2026.10},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/NodeNovaOrg/bidlens-open-data},
  license   = {CC-BY-4.0}
}