hemanthreddy901/nadora-global-industries

NADORA Global Industries

A synthetic multinational, built to be developed against rather than demonstrated with.

One fictional company — $5.20bn revenue, $716m EBITDA, 24,000 employees, 18 countries, 35 legal entities, five business units — traded daily from January 2022 to December 2026 and rendered at six fidelities, from a 3 MB unit-test fixture to a 5 GB full-scale corpus.

38,964,663 rows · 11 GB · 2,319 verification assertions, all passing.

100% synthetic. No real company, no personal data.

---

The six levels

	Purpose	Tables	Rows	Checks	FKs	Anomalies	Defects	Docs	Size
L0 Fixture	unit tests and CI; builds in 2s	40	23,181	262/262	73	—	—	—	3 MB
L1 Clean	does ingestion and the workflow work?	40	124,003	262/262	73	1	—	—	16 MB
L2 Messy	can it reconcile inconsistent data?	49	1,355,680	337/337	102	14	34,698	—	417 MB
L3 Cross-functional	can it connect departments?	65	5,914,640	398/398	117	23	217,702	1,161	1.6 GB
L4 Enterprise	can it understand a multinational?	73	12,964,283	430/430	125	25	529,445	3,816	3.6 GB
L5 Extreme	can it find relationships across millions of records?	73	18,582,876	430/430	125	25	846,703	5,656	5.1 GB

Each level adds capability, not only volume. L2 turns on multi-currency, the dirt engine and Excel workbooks. L3 adds the remaining departments, five years, matrix reporting and the document corpus. L4 adds 18 countries, 35 entities, intercompany trade, fiscal calendars, a $450m acquisition and real PDF/Word/PowerPoint. L5 runs the full transaction density.

Start with L1

16 MB, 40 tables, 73 foreign keys, a complete company. If your pipeline handles L1 it handles the shape of all of them — see the nesting contract below. Move up when entity resolution matters (L2), when departments have to talk to each other (L3), when you need consolidation and binary documents (L4), and when you want scale (L5).

---

What is in a level

L<n>/
  clean/                   every table as CSV and Parquet, pristine
  dirty/                   the same tables with located defects injected
  workbooks/               16 .xlsx across 4 upload bundles x 4 difficulty tiers
  documents/               meetings, emails, memos, functional reports, board
                           packs, plant reports, supplier scorecards, account
                           reviews, financial statements, audit reports
  ground_truth.json        the declared model, measured stats, the causal graph,
                           the planted anomalies, the known contradictions
  verification.csv         every assertion, run on clean data before dirtying
  dirt_manifest.json       every defect located, plus the entity-resolution key
  pathologies.json         which spreadsheet corruption hit which sheet
  document_evidence.json   facts that exist in a document and in no table
  questions.jsonl          scoreable questions with typed expected answers
  manifest.json            build config, per-table counts, timings

File-by-file

File	What it is	Use it for
clean/*.csv	40–73 tables, pristine, UTF-8, plain header row	The reference truth. pd.read_csv(path) works with default settings on every table.
clean/*.parquet	The same rows, typed and compressed (1.5 GB vs 8.2 GB)	Fast loading and column pruning. Identical content to the CSV.
dirty/*	The same tables after the dirt engine	Point your pipeline here. This is what an enterprise actually sends. Tables the dirt engine does not touch are copied byte-identical.
workbooks/*.xlsx	Four bundles (finance/HR, sales, product/supply, marketing) at four tiers	Tier 0 is a clean export. Tier 3 has title rows, merged headers, a notes column, a total row that is not data, blank separators and numbers stored as text.
documents/	Markdown, text, PDF, Word, PowerPoint	Some facts appear in exactly one document and in no table. A platform that never reads the documents cannot find them.
ground_truth.json	The answer key	Scoring. Keep it out of your ingestion path.
verification.csv	Every assertion with measured vs expected	Proof the level is internally consistent before it was dirtied.
questions.jsonl	17 questions at L2, 28 at L3+	Machine-checkable evaluation. See the scoring section.

---

The nesting contract

The levels are nested supersets. Every identifier present at level N is present at level N+1, describing the same entity with the same attribute values. CUST00307 is the same customer, with the same segment, terms and over-generous discount, at every level that contains it. Order SO00000003 has the same customer, date, lines and value everywhere.

So code written against the 16 MB level runs unmodified against the 5 GB one. You change a path, not your parser.

Verified on the published files for all eight populations, including 820,000 sales orders and 540,000 purchase orders.

Two columns are exempt, and the exemption is recorded in ground_truth.json: employees.site_id (which of eighty buildings a person occupies depends on how much of the footprint the level builds) and employees.dotted_line_manager_id (matrix reporting switches on at L3).

How transaction ids nest, and why it is not obvious

Every order is built from its own index, drawn from the full universe of customers, products and months; a level keeps the ones whose participants it contains. The subtlety is where to stop. Stopping at a target row count breaks the guarantee, because a larger level accepts a higher share of the stream and so reaches its quota after scanning fewer indices than the level below — orders the smaller level contained would never be looked at. Each level therefore declares a scan depth, the depth never decreases, and the row count is whatever falls out.

---

How it was made

A monthly simulation runs once; the six levels are renderings of it.

Supplier quality degrades. Bad material stops production lines. Stoppages force overtime. Sustained overtime drives people out. Leavers take experience with them. Inexperienced replacements make more mistakes. Those mistakes become warranty claims four to ten months later and support tickets three months later. All of it lands in the ledger as variances.

Every one of those is a line of arithmetic in the generator. None of it is written down anywhere in the data. The only way to recover a relationship is to correlate across departments and respect the lag — which is why correlating within a single month finds nothing.

Transactions are sampled so they aggregate back to the monthly panel exactly. That is what lets the verification suite assert that GL revenue equals the sum of non-cancelled sales line values to the cent ($0.0000 gap on $28.15bn across 2,916,257 journals, none unbalanced). That internal consistency is what makes a wrong answer detectable at all.

The five-year arc

2022	normal growth
2023	raw-material inflation, energy spike, tight labour market
2024	long-haul supplier disruption, port congestion, new product launch
2025	$450m acquisition, integration problems, ERP outage
2026	demand slowdown, restructuring

---

The evaluation layer

25 planted anomalies (6 easy, 9 moderate, 8 hard, 2 very hard). A supplier that reprices 34% mid-contract. A plant whose scrap runs at twice the network. A manager whose team leaves fastest of any with fifteen or more reports. A top-ten customer that is loss-making once service cost is loaded (−13.8% net, against +44% for the next-worst). Duplicate invoices, split purchase orders, amounts clustered under authority limits, bank details changed before a large payment. Each records how it should be found and how hard it should be. Four are downstream symptoms of another, so a platform reporting five independent problems where there are two has not understood the data.

11 known contradictions. Six metrics reported by two functions on two definitions; the five-number forecast ladder (pipeline / sales commit / finance plan / operations capacity / actual); the three-number savings gap (procurement claims, finance recognises less, operations realises less again); intercompany legs that do not match. Each says who is right, why, and how to reconcile. None is random.

20 causal edges with lags and strengths, and the tables each is visible in.

Scoring

questions.jsonl holds typed, machine-checkable answers — entity sets, ranked lists, causal chains, values with tolerances.

python -m generator.evallayer.score L3 my_answers.json

Entity sets score by F1, ranked lists by rank-biased overlap, causal chains by set overlap plus ordering. Unanswered scores zero, because a system that declines to answer has not answered. Reporting everything as a problem scores near zero too.

examples/baseline_solver.py is a deliberately simple pandas baseline that reads only clean/ and never opens the answer key. It scores about 0.18. That is the intended result: the easy questions fall to straightforward analysis, and the ones that matter — why EBITDA margin declined, what caused warranty costs to rise, which risks appear only in the meeting notes — do not.

---

Loading

import pandas as pd, json
from huggingface_hub import snapshot_download

# just one level, rather than all 11 GB
path = snapshot_download("hemanthreddy901/nadora-global-industries",
                         repo_type="dataset", allow_patterns="L1/*")

orders = pd.read_parquet(f"{path}/L1/clean/sales_orders.parquet")
dirty  = pd.read_parquet(f"{path}/L1/dirty/customers.parquet")

gt = json.load(open(f"{path}/L1/ground_truth.json"))
gt["relationships"]          # 73 foreign keys with cardinality
gt["causal_graph"]["edges"]  # 20 labelled relationships with lags

Plain CSV works everywhere with no special settings:

pd.read_csv(f"{path}/L1/clean/sales_orders.csv")

---

Reproducing it

The generator is included.

python3 -m venv .venv
.venv/bin/pip install pandas numpy pyarrow openpyxl reportlab python-docx python-pptx pytest
.venv/bin/python -m generator.build L1     # one level
.venv/bin/python -m generator.build all    # all six, about 70 minutes
.venv/bin/python -m pytest tests/ -q       # 138 tests, about a minute

One seed (20260915) drives everything; rebuilding is deterministic and a test enforces it. The suite covers the declared model, the nesting contract (in the builders and against the emitted Parquet), the ledger tie-out, the organisation's acyclicity, that every planted anomaly is anchored on an entity that exists at the level it claims, that every detection route names a real table, that no conclusion leaked into the data, and that the scorer returns 1.0 for correct answers and near-zero for crying wolf.

---

A warning about the answer key

ground_truth.json, dirt_manifest.json, document_evidence.json and questions.jsonl are the answers. Keep them out of whatever you point your ingestion at. The data itself is clean of them — a test checks that no anomaly id, root cause or conclusion appears in any column of any table.

---

Known limits

• One volume target is not met, deliberately. inventory_snapshots holds 749,492 rows at L5 against the 1,000,000+ originally specified. It came down when a modelling error was fixed: each SKU had been stocked at thirteen warehouses, where a real network holds two or three. Row count was traded for a network that resembles reality; raising it is a config change, not new code. Every other volume target is met or exceeded — 820,000 sales orders, 540,000 purchase orders, 599,974 service tickets, 9,329,369 journal lines.
• Lower levels are extracts, not miniatures. L1 holds the largest 200 customers, so its concentration and margins differ from the group. Absolute financials only match the spec at L5; manifest.json records each level's configuration.
• Column statistics for gl_transactions at L3–L5 are computed from the first 12,000 rows, because the ledger is streamed to disk during the build rather than held in memory. ground_truth.json says so explicitly for that table; the row count is the true one.