Add datasets query_imdb and query_krama

[Ruiying-Ma]

Datasets Added

imdb

• Domain: Movie/entertainment industry (IMDB JOB benchmark)
• Databases: PostgreSQL (movies_db) + SQLite (people.sqlite)
• Properties:
  • Multi-DB integration: queries span PostgreSQL and SQLite
  • Ill-formatted join keys: identifier columns use non-standard string encodings (e.g. tt0000042, nm001, InfT~~7); agents must strip prefixes and leading zeros before
    joining

krama

• Domain: Multi-domain scientific data (wildfire, environment, legal, biomedical, astronomy, archeology)
• Databases: MongoDB (domain_docs.bson for csv/txt files) + SQLite (us_geo.db for gpkg files and a synthetic beach_water_temperature table) + SQLite (domain_assets.db for other files)
• Properties:
  • Multi-DB integration: queries span MongoDB, SQLite, and heterogeneous flat files
  • Ill-formatted join keys: beach codes: plain integers vs 'COM-XXX' prefix; UCEC patient IDs: 'S001' vs 's_1'
  • Unstructured-text-transformation: GACC / location / doc info merged into NL description field

Each dataset has 10 queries with ground-truth CSVs and validate.py scripts.

---

Validation Results (Claude Sonnet 4.6, 1 run/query)

Dataset	Accuracy	Timeouts	Avg Latency (successful runs)	Min	Max
imdb	5/10 (50%)	3	929.6s (~15.5 min)	198s	2214s
krama	9/10 (90%)	0	129.8s (~2.2 min)	72s	236s
Overall	14/20 (70%)	3	—	—	—