hemanthreddy901/enterprise-corpus

Enterprise Corpus — tabular relationship inference & forecasting

Twelve synthetic corpora modelling a multinational business, generated from a declared relational model so the correct answer is known by construction and proven by assertion.

Built to benchmark systems that ingest tabular uploads, infer how the tables relate, produce insights, and forecast.

Corpora	12
Rows	16,476,554
Tables per corpus	22
Foreign keys per corpus	36
Excel workbooks	132
Injected defects	5,752,430
Verification	124/124 checks passing on every corpus

Everything here is 100% synthetic. No personal data, no real company data.

---

Why this exists

A folder of realistic spreadsheets gives you demos. It does not tell you whether an engine is right, because nobody knows the correct answer.

This corpus works backwards: it declares the relational model — primary keys, foreign keys, cardinalities, grains — generates coherent business data from it, verifies those declarations against the data actually produced, and only then degrades everything into the kind of Excel an enterprise really uploads.

---

What the data represents

A mid-size multinational that buys, stocks, sells and accounts for physical goods, traded daily for three years (2022-01-01 → 2024-12-31) across 13 countries and 11 currencies.

Trade customers place orders with a named sales rep; orders ship from a distribution centre, are invoiced in local currency, and are delivered, shipped, processing or cancelled. Procurement raises POs on approved suppliers with real lead times. Stock is counted at every warehouse each month end. Marketing runs campaigns against selected products. Finance posts it all to a chart of accounts.

It reconciles: GL revenue accounts tie back to sales line values with a 0.00% gap. That internal consistency is what makes wrong answers detectable.

Four industry variants share the structure and differ in vocabulary and seasonality: consumer_goods, industrial, pharma_distribution, apparel_retail.

---

Quick start

import pandas as pd, json
from huggingface_hub import snapshot_download

path = snapshot_download("hemanthreddy901/enterprise-corpus", repo_type="dataset")

gt = json.load(open(f"{path}/consumer_goods_medium_mixed_dirty/ground_truth.json"))
edges = [(r["from_table"], r["from_columns"], r["to_table"],
          r["to_columns"], r["cardinality"]) for r in gt["relationships"]]
# -> 36 labelled foreign keys

lines = pd.read_parquet(
    f"{path}/consumer_goods_medium_mixed_dirty/dirty/sales_order_lines.parquet")

Grab a single corpus instead of all 2.5 GB:

snapshot_download("hemanthreddy901/enterprise-corpus", repo_type="dataset",
                  allow_patterns="consumer_goods_medium_mixed_dirty/*")

No account or token is needed to download.

---

What ships with every corpus

clean/              22 tables as CSV + Parquet, pristine
dirty/              the same tables with defects injected
workbooks/          .xlsx across 4 upload bundles x N difficulty tiers
ground_truth.json   36 FKs with cardinality, PKs, grains, roles, column stats,
                    the true generating process, column_name_map, difficulty score
verification.csv    124 assertions, run on clean data before dirtying
dirt_manifest.json  every defect located, with a how_to_score note
pathologies.json    which Excel corruption hit which sheet

Ground truth is verified before dirt is injected, so the labels are proven against pristine data while the workbooks are built from the dirty version.

---

The 22 tables

Table	Role	Primary key	Represents
regions	dimension	region_id	EMEA / AMER / APAC reporting regions
countries	dimension	country_id	Countries of operation
currencies	dimension	currency_code	Trading currencies — natural string key
fx_rates	rate	currency_code + rate_date	Daily USD conversion rate
date_calendar	calendar	date_key	Calendar + fiscal attributes, FY starts April
product_categories	hierarchy	category_id	Two-level taxonomy in one table
products	dimension	product_id	Sellable range with lifecycle dates
customers	hierarchy	customer_id	Trade accounts with corporate rollups
suppliers	dimension	supplier_id	Vendors with lead times and terms
product_suppliers	bridge	product_id + supplier_id	True many-to-many
warehouses	dimension	warehouse_id	Distribution centres
employees	hierarchy	employee_id	Staff across 8 departments, manager chains
gl_accounts	hierarchy	account_code	Chart of accounts with parent rollups
marketing_campaigns	dimension	campaign_id	Campaigns with budget and planned lift
campaign_products	bridge	campaign_id + product_id	True many-to-many
sales_orders	fact	order_id	Order headers
sales_order_lines	fact_line	line_id	Line detail — finest transactional grain
sales_returns	fact	return_id	Returns against delivered orders
purchase_orders	fact	po_id	Inbound POs
purchase_order_lines	fact_line	po_line_id	PO line detail
inventory_snapshots	snapshot	warehouse_id + product_id + snapshot_date	Month-end stock — semi-additive
gl_transactions	fact	gl_id	Monthly postings by account and country

---

Structures that break naive engines

Structure	Where	The failure it catches
True M:N bridge	product_suppliers	Row inflation 2.06×; SUM(unit_cost) overstated 108%
Semi-additive snapshot	inventory_snapshots	136M units summed across dates vs the correct 3.7M
Self-referencing hierarchy	4 tables	Parent and child both summed — silent double-count
Role-specific FK	sales_orders.sales_rep_id	Targets a subset; 276 of 400 employees never referenced
Composite reference	sales_returns → lines	Needs two columns matched together
Redundant join path	sales_orders.country_id	One relationship reported as two independent facts
As-of join	fx_rates by order_date	A dated lookup treated as an equi-join

Five further relationships are declared in semantic_relationships that no key-matching algorithm can find — procurement lead/lag, GL reconciliation, campaign causality, FX conversion, denormalisation. Score those separately.

---

Difficulty controls

Column naming

With generated names, 31 of 36 foreign keys are solvable by exact column-name match alone — a twenty-line matcher scores 86% with no inference. Corpora suffixed _mixed disguise columns per table, as four files from four systems:

clean       customer_id      order_date       net_amount
sap         KUNNR            AUDAT            NETWR
legacy      CUST_NO          ORD_DT           NET_AMT
verbose     Customer Account Number            Net Value
bilingual   Kundennummer     Bestelldatum     Nettobetrag
opaque      Col1             FIELD_5          F2

That drops name-solvability to 25%. ground_truth.json → difficulty records the figure, so the benchmark reports its own honesty.

Eight defect classes (_dirty corpora)

Defect	Correct handling
orphan_fk — 3.0% point at non-existent IDs	Report the relationship anyway
duplicate_entity — Clearwater Supply Inc / Inc.	Merge; clusters supplied
missing_not_at_random — ship_date, credit_limit	Do not impute
blank_means_zero — discount_pct	Fill with zero
impossible_value — negative qty, dates in 2099/1900	Flag as invalid
unit_inconsistency — 8% of weight_kg in grams	Detect bimodal distribution
duplicate_rows — re-uploaded file signature	De-duplicate
free_text_column — notes embedding a real po_number	Surface the text-only link

The imputation pairing is deliberate: credit_limit blanks must not be filled, discount_pct blanks must be. Same shape, opposite answers.

Four Excel tiers

0 pristine · 1 system export (title block, header on row 5, numbers as text, mixed date formats) · 2 analyst-touched (merged two-row headers, embedded subtotals, month-tab splits needing a union, pivoted P&L) · 3 in the wild (hidden columns, #REF!, CUS-000706 / cus-000477, meaning in fill colour, near-duplicate customers FINAL sheets)

---

Forecast-grade time series

Sales come from a known process, recorded in data_generating_process, so forecast quality can be scored against the true signal rather than only a holdout: 8.5%/yr trend, weekday shape (Mon 18.9% of revenue → Sun 4.9%), holiday effects (Black Friday 1.65×), product lifecycle, campaign lift, and an 8-week supply shock at 55% through the history.

Month-of-year index (1.00 = average month):

consumer_goods  1:0.71 ... 11:1.45 12:1.37
industrial      1:0.82 ...  7:0.73 12:1.32
pharma          1:1.24 ...  7:0.73 12:1.25
apparel         1:0.64 ... 11:1.44 12:1.41

Pharma peaks in January while apparel bottoms out there at 0.64. A forecaster that learned "Q4 high, Q1 low" from one vertical gets pharma backwards.

A built-in trap: the naive campaign-lift estimate reads ~51% against a true injected lift of ~25%, because the window carries trend and seasonality too. Recovering the real figure needs a control group or a seasonality-adjusted model.

---

The twelve corpora

Hard mode — mixed naming + injected dirt

Corpus	Rows	Workbooks	Defects
industrial_large_mixed_dirty	12,735,187	8	4,721,338
consumer_goods_medium_mixed_dirty	1,659,139	16	667,136
apparel_retail_small_mixed_dirty	276,961	16	116,788
consumer_goods_small_mixed_dirty	273,611	—	114,690
pharma_distribution_small_mixed_dirty	269,998	16	112,176
consumer_goods_tiny_mixed_dirty	59,482	4	20,302

Clean baseline — the control group

Corpus	Rows	Workbooks
apparel_retail_small	276,961	16
consumer_goods_small	273,611	16
pharma_distribution_small	269,998	16
industrial_small	261,377	16
consumer_goods_tiny	59,982	8
apparel_retail_tiny	60,247	—

Same seed, same row counts — the only variables between a pair are naming and dirt, so the difficulty delta is cleanly attributable. Two corpora have no workbooks (they were --no-excel runs) and remain fully usable for inference testing.

Start with consumer_goods_medium_mixed_dirty, paired with consumer_goods_small as the control.

---

Suggested metrics

Structure — FK-graph precision/recall · PK and grain detection · cardinality classification, tracking 1:N mistaken for N:M separately because it silently corrupts every downstream number · union-vs-join classification on month tabs.

Interaction — number of confirmation questions asked (budget: ≤7 before users abandon) · accuracy gained per question.

Insight & forecast — recovery of the known DGP components · error vs a seasonal-naive baseline · does it detect the supply shock · does it report the naive 51% campaign lift or the true 25%.

Operational — determinism · cost and latency per workbook · accuracy per tier and per pathology.

The Valentine benchmark (ICDE 2021) formalised this sub-problem and introduced recall @ ground truth for ranking quality — worth adopting rather than reinventing.

---

Reproducibility

Every corpus is byte-reproducible from its seed — verified: all 22 tables hash-identical on regeneration. The generator is ~360 KB of Python, so regenerating is often cheaper than downloading.

python -m gen.run --industry consumer_goods --scale medium --naming mixed --dirt

Scales: tiny 60k rows / 5s · small 274k / 48s · medium 1.66M / 100s · large 12.7M / 204s.

---

Limitations

• One schema in four costumes. Real diversity means different people's design decisions. Pair this with real relational databases for schema variety.
• Unknown unknowns. A generator only injects defects someone thought of. Real data carries mid-history rule changes and half-finished migrations.
• Overfitting risk. Tuning against a generator is overfitting with extra steps. Hold out real workbooks the engine is never tuned on.