minhaozhang/BIRD_Obfuscation

BIRD Obfuscation: PostgreSQL dumps

A cleaned, contamination-resistant, adversarial rebuild of the BIRD Text-to-SQL benchmark, curated as the substrate for a semantic-layer agent evaluation rather than for one-shot Text-to-SQL.

Code, gold SQL, manifests and full documentation: github.com/Minhao-Zhang/BIRD-Obfuscation

This Hub repo holds only the four database dumps. The questions, gold SQL, rename map, trap manifests and per-question gold-quality provenance are git-tracked in that repository under eval_dataset/. You need both halves.

What is in the dumps

Four PostgreSQL custom-format dumps (pg_dump -Fc, zstd-compressed), one per arm. Each dump is a whole bird database containing all 69 obfuscated BIRD databases as 69 schemas plus public. Real rows, columns and tables are byte-identical across all four; only identifiers and the presence of decoy objects differ.

Produced with PostgreSQL 18.4. Restore into PostgreSQL 18 or newer with pg_restore.

file	arm	port	identifiers	decoys/traps	size	TOC entries
pg_base.dump	base	5432	original English	none	2.93 GB	1591
pg_rename.dump	rename	5433	renamed (target language)	none	2.93 GB	1591
pg_decoy.dump	decoy	5434	original English	corrupted traps	3.13 GB	1915
pg_rename_decoy.dump	rename+decoy	5435	renamed	corrupted traps	3.13 GB	1915

Identifier rename. Table and column names are translated into one of five languages, roughly 14 databases each: English (identity, kept as a noise-floor control), French, German, Spanish and Mandarin Pinyin. The English slot exists so a subset of databases has a guaranteed near-zero rename effect to measure against.

Corrupted traps. The two decoy arms add 1,486 evil-twin columns and 162 cloned tables holding subtly wrong copies of real data under plausible synonym names, aimed at agents that explore a schema by running queries. Every trap is strictly additive, which is what keeps the ground-truth task provably intact: obfuscated gold stays execution-equivalent to the validated original (R0==R1 against SQLite, R1==R2 across instances) for every retained question.

The TOC delta is a useful sanity check on that claim: 1915 minus 1591 is 324, which is exactly the 162 clone tables times two entries each (TABLE plus TABLE DATA). Evil-twin columns are ALTERs on existing tables and so add no TOC entries. Nothing else was injected.

The paired question set

Numbers below describe the gold data in the GitHub repo, not these dumps.

Databases evaluated	57 of the 69 present in the dumps
Questions	6,743 (5,392 train / 1,351 test)
Per-database test fraction	uniform 19.4% to 20.6%
Per-database corpus size	61 to 383 questions
Cross-split duplicate leakage	0.22% of the test set

BIRD's own gold SQL is substantially and systematically mis-annotated. Published audits report 49% to 61% error rates depending on the split. 2,739 questions were removed from this dataset by joining against the official bird_sql_dev_20251106 (corrected dev) and bird23-train-filtered (filtered train) releases, after which 11 databases fell below the 60-question floor and were dropped. A later pass removed 127 duplicate questions, taking cross-split leakage from 3.6% of the test set to 0.22%. Per-question provenance for all 10,164 pre-purge questions, including the corrected gold SQL for every changed row, ships as gold_quality_flags.jsonl, and the collapsed duplicate clusters as dedupe_clusters.json. Method, evidence and citations: gold-quality-audit.md.

The dumps carry 69 schemas; the evaluation covers 57. The 12 databases that fell below the 60-question floor were not removed from the dumps, so they are still present, fully obfuscated and trap-injected, but no question, gold SQL or result hash references them: app_store, bike_share_1, california_schools, college_completion, computer_student, cookbook, debit_card_specializing, financial, music_platform_2, retail_world, sales_in_weather, software_company.

For an agent over a pooled schema lake they are either extra distractors that make routing genuinely harder, or wasted exploration budget. Either reading is defensible, but it sets the routing_recall denominator, so state which one you used. Drive an evaluation from evaluated_dbs.json (57); use retained_dbs.json (69) only to reason about what a probing agent can see.

Verify

442C830F0FF598FAEB10AEB30291FE5A628E10079A9DCC526F32F0C72BBD274D  pg_base.dump
4A89D6853562D29825244C416877734C038E75B7A47D31287A2A87F39D24AE36  pg_rename.dump
791D5FF65CD457095E3E7FF4484F693077AA45E42F4F506C6BE09AE9DD9B0C26  pg_decoy.dump
43282AA7C497B66D0B1CCE0EB9BB574CFB7925821AE4179354221524B6F4CA28  pg_rename_decoy.dump

sha256sum -c SHA256SUMS.txt on Linux or macOS, Get-FileHash <file> -Algorithm SHA256 in PowerShell.

Restore

Each dump restores into a fresh database. --no-owner drops the dependency on the original bird role so any superuser can restore it, and -j restores in parallel.

Into an existing server, one target database per arm you need:

createdb bird_base
pg_restore -d bird_base --no-owner --no-privileges -j 4 pg_base.dump

Into a fresh container, mirroring the repo's Compose setup:

docker run -d --name pg_base -e POSTGRES_USER=bird -e POSTGRES_PASSWORD=bird \
  -e POSTGRES_DB=bird -p 5432:5432 postgres:18
docker cp pg_base.dump pg_base:/tmp/pg_base.dump
docker exec pg_base pg_restore -U bird -d bird --no-owner -j 4 /tmp/pg_base.dump

Post-restore sanity check: a clean arm has 569 base tables across 70 schemas, and a decoy arm has 731 (569 real plus the 162 clone tables).

SELECT count(DISTINCT table_schema), count(*) FROM information_schema.tables
WHERE table_type = 'BASE TABLE' AND table_schema NOT IN ('pg_catalog', 'information_schema');

Notes

• These are logical dumps, portable and version-flexible, not physical volume copies, so they restore cleanly on any OS or architecture running PostgreSQL 18 or newer.
• The dumps carry no indexes, primary keys or foreign keys. That is deliberate, not an export artifact: the benchmark withholds an explicit foreign-key catalogue so an agent has to infer table relationships from column names, values and the question/SQL pairs it has seen. It also makes restore fast.
• Full setup, obfuscation methodology, evaluation design and known limitations are documented in the GitHub repository. Read limitations.md before citing any number.

License and attribution

CC BY-SA 4.0. Share and adapt with credit, under the same license. This dataset is a derivative of the BIRD benchmark; please credit BIRD as the upstream source, and the birdsql corrected releases linked above if you rely on the gold-quality filtering.