Project Awesome project awesome
♥

Poland Public Tenders Dataset (BZP + TED)

Open dataset of Polish public procurement notices [...]

Package GitHub

Polish Public Tenders Dataset

DOI Kaggle Data license: CC BY 4.0 Code license: MIT Release Records Anonymization Python Atlas Przetargów

Open dataset of Polish public procurement notices aggregated from official sources: BZP (Biuletyn Zamówień Publicznych, Polish national portal) and TED (Tenders Electronic Daily, EU-wide database).

Maintained by Atlas Przetargów — the Polish public procurement search and analytics platform.

  • Coverage: 2024 — present (earlier years planned)
  • Records: ~1.4 million tender notices, ~100 000 buyer profiles, ~80 000 contractor profiles
  • Update cadence: quarterly (full dump per release)
  • Format: CSV (UTF-8) + Parquet
  • License: CC BY 4.0 for data, MIT for code
  • Privacy: Natural-person contractors (sole proprietors) are anonymized — see Privacy & anonymization below.

Contents

data/
  tenders_YYYY.csv          # Yearly tender dumps
  tenders_YYYY.parquet      # Same, columnar format (recommended for analytics)
  buyers.csv                # Aggregated buyer profiles (by NIP)
  contractors.csv           # Aggregated contractor profiles (by NIP, anonymized for natural persons)
  city_cache.csv            # City name → lat/lng/province mapping
schema/
  tenders.md                # Column descriptions for tenders
  buyers.md
  contractors.md
export.py                   # Regenerate data from Atlas Przetargów database
audit_pii.py                # PII audit report generator
pii_utils.py                # PII detection + anonymization (used by export.py)
publish.sh                  # Publish to GitHub + Kaggle + Zenodo
LICENSE-DATA                # CC BY 4.0 (for data)
LICENSE-CODE                # MIT (for code)

Quick start

For a full runnable tour — including a bar chart of top CPV divisions, a Herfindahl-Hirschman market-concentration calculation, and a DuckDB-over-HTTP example — see notebooks/01_getting_started.ipynb. Rendered directly by GitHub, no local setup needed to preview.

Pandas

import pandas as pd

tenders = pd.read_parquet("data/tenders_2024.parquet")
print(f"{len(tenders):,} tenders in 2024")
print(tenders.groupby("province")["estimated_value"].sum().sort_values(ascending=False).head())

SQL (DuckDB)

SELECT province, COUNT(*) AS n, SUM(estimated_value) AS total_value
FROM 'data/tenders_2024.parquet'
WHERE notice_type LIKE 'Contract%'
GROUP BY province
ORDER BY total_value DESC;

Schema highlights

Column Type Description
id string BZP or TED notice number (primary key)
title string Tender title
buyer string Contracting authority name
buyer_nip string Polish tax ID (10 digits)
city, province string Location; province uses PLxx codes (NUTS-2 compatible)
cpv_code string Comma-separated CPV codes (EU procurement vocabulary)
notice_type string ContractNotice, TenderResultNotice, cn-standard (TED), etc.
order_type string Roboty budowlane / Dostawy / Usługi
date date Publication date
submitting_offers_date timestamp Deadline for offer submission
estimated_value float Estimated contract value (in currency)
currency string Usually PLN or EUR (TED)
source string bzp or ted
is_duplicate bool TED duplicates of BZP entries (filter these out for deduplicated analyses)
contractor_name, contractor_national_id string Winner (only on TenderResultNotice)
latitude, longitude float Geocoded coordinates of buyer's city

Full schema: schema/tenders.md.

Typical use cases

  • Market analysis — which provinces/cities issue most tenders, by sector (CPV)
  • Buyer/contractor profiling — who wins what, from whom, for how much
  • Academic research — public procurement economics, competition, pricing
  • NLP — Polish-language tender titles and specifications (large corpus)
  • ML baselines — classification (CPV prediction), value estimation, winner prediction

Attribution

If you use this dataset, please cite it. A machine-readable CITATION.cff is included so GitHub, Zenodo, and reference managers (Zotero, Mendeley) can import the citation automatically.

Human-readable citation:

Atlas Przetargów (2026). Polish Public Tenders Dataset (BZP + TED), version 2026.Q2. https://github.com/atlasprzetargow/polish-tenders-dataset

Or, if using the Zenodo-archived version (stable DOI, preferred for academic citations):

Atlas Przetargów (2026). Polish Public Tenders Dataset (BZP + TED) — 2026.Q2. Zenodo. https://doi.org/10.5281/zenodo.19634050

Data sources

Data is aggregated from official public sources:

  • BZP (ezamowienia.gov.pl) — Polish national procurement portal, run by UZP
  • TED (ted.europa.eu) — EU-wide procurement database, run by the Publications Office of the EU

This dataset adds on top of the raw sources:

  • Deduplication — TED notices that duplicate BZP entries are flagged (is_duplicate = true)
  • Geocoding — cities resolved to lat/lng
  • Normalization — province codes standardized to PLxx, NIPs cleaned
  • Aggregation — per-buyer and per-contractor profiles with totals and trends
  • Anonymization — natural-person contractors (CEIDG sole proprietors, PESEL holders) are anonymized; see below

Privacy & anonymization

Polish public procurement contract-award data is, by law, public: Article 269 of the Polish Public Procurement Act (PZP) mandates publication of contractor names and NIPs in the BZP bulletin, and the EU Open Data Directive 2019/1024 actively encourages re-use of procurement data. Aggregator platforms (eGospodarka, Oferent, Bazhub, etc.) have been republishing this data for over a decade without anonymization; the Polish Data Protection Authority (UODO) has taken no enforcement action. No anonymization is legally required, either for buyers or contractors.

That said, bulk redistribution of 1.4M records under a CC BY license — downloadable by any crawler — differs qualitatively from per-query lookups on a website. As a precautionary measure, we additionally anonymize contractor rows where our heuristic classifies the winner as a natural person. This is optional, not GDPR-mandated. Rules applied:

  • contractor_national_id is 11 digits long (PESEL — Polish personal identifier; companies use 10-digit NIPs)
  • contractor_name contains explicit markers such as "osoba fizyczna", "jednoosobowa działalność", "prowadzący działalność", or CEIDG
  • contractor_name ends with a "Imię Nazwisko" pattern AND the first-name token matches a curated list of ~250 Polish given names (eliminates false positives for foreign corporate names)

When any of the above match, the row is transformed as follows:

Field Before After
contractor_name "ABC Firma Usługowa Jan Kowalski" "[Osoba fizyczna]"
contractor_national_id "9876543210" "anon-a1b2c3d4e5" (stable SHA-256 hash, salted)
contractor_city "Kraków" kept — city-level aggregation still works
contractor_province "PL21" kept
contractors (JSON) per-entry with name+ID per-entry anonymized with the same rules

The hash is stable across releases (same input → same output) but irreversible without the salt. This preserves cross-year joins and contractor-level analytics while removing identifying information.

Buyer side is NOT anonymized. Polish public-procurement buyers are, by law, public bodies or publicly-registered entities; there are zero PESEL-length IDs in the buyer_nip column.

Coverage statistics (2024 sample): 59,472 of 238,103 rows with a contractor name (~25%) are anonymized under the above rules. The anonymization code is open-source under MIT in pii_utils.py and audit_pii.py. If you believe a row has been mis-classified (false positive or negative), please open an issue with the tender id.

Regenerating the dataset

# From the Atlas Przetargów monorepo
cd open-data
python export.py --years 2020-2025 --output ./data

Requires:

  • Python 3.11+, pandas, pyarrow, sqlalchemy, psycopg2-binary
  • Access to the Atlas Przetargów database (env var DATABASE_URL)

Public contributors cannot regenerate — the full DB is not public. But the output CSVs are open under CC BY 4.0.

Contributing

  • Issues — found a data quality problem? Open an issue with the tender id.
  • Schema suggestions — PRs welcome on schema/ docs.
  • Code — export.py and publish.sh are MIT-licensed; PRs for better formatting, additional derived columns, etc. welcome.

Mirrors and citations

The same dataset is mirrored on Zenodo (for stable DOI / academic citation) and Kaggle (for ML / notebook discoverability):

Related

Changelog

  • v2026.Q2 (2026-04-17) — Initial public release. Coverage 2024 — present (~1.4M tender notices). PII anonymization enabled for all natural-person contractors.
Back to Public Datasets