vasilisplavos/unboxgov-khmdhs

unboxgov — KHMDHS contracts

Structured Greek public-procurement contracts from the KHMDHS OpenData API, plus anomaly-detection flags surfaced by the unboxgov pipeline.

What's in here

• metadata/contracts.parquet, contract_items.parquet, flags.parquet, orgs.parquet — cumulative analytical tables (see schema.json for columns). contracts ↔ contract_items join on contract_id; flags are the detected patterns; orgs is per-entity (AFM) aggregates.
• metadata/raw/<YYYY>.parquet — the full original KHMDHS payload per contract ((referenceNumber, raw), raw is the JSON record as text), partitioned by submission-year (unknown.parquet holds rows with no submission date). Complete — every contract, not just flagged ones.
• contracts.sqlite — the same full payloads as a point-lookup key-value store: table contracts(referenceNumber TEXT PRIMARY KEY, doc TEXT), where doc is the JSON envelope {"schemaVersion", "referenceNumber", "khmdhsPayload"}. Built for fetching one contract by id cheaply (indexed seek) — including from a browser over HTTP Range, with no full-file download.
• reports/<YYYY_MM>/ — generated Greek-language summaries.

Using the data

Analytical tables (pandas / DuckDB / the 🤗 datasets loader):

import duckdb
BASE = "https://huggingface.co/datasets/vasilisplavos/unboxgov-khmdhs/resolve/main"
duckdb.sql(f"SELECT rule, COUNT(*) FROM '{BASE}/metadata/flags.parquet' GROUP BY 1 ORDER BY 2 DESC")

One contract by reference number — from contracts.sqlite (an ordinary SQLite file):

import sqlite3, json
from huggingface_hub import hf_hub_download

path = hf_hub_download("vasilisplavos/unboxgov-khmdhs", "contracts.sqlite", repo_type="dataset")
con = sqlite3.connect(path)
(doc,) = con.execute(
    "SELECT doc FROM contracts WHERE referenceNumber = ?", ("25SYMV017220040",)
).fetchone()
print(json.loads(doc)["khmdhsPayload"])

The referenceNumber primary key is indexed, so single-id lookups are fast and need only a few pages of the file — which is why the website queries contracts.sqlite directly in the browser (via sql.js-httpvfs, HTTP Range) instead of downloading it. For bulk work over all original payloads, prefer the metadata/raw/<YYYY>.parquet partitions (columnar, compressed).

Disclaimer

Flags are indications worth reviewing, not findings of wrongdoing. Each flag surfaces a pattern; verifying it requires a human reviewing the underlying contracts.