scm-sql / docs /STATISTICS.md
AniruddhaAI's picture
SCM-SQL v1.0 - 500 pairs / 556 turn-level trials / 6 complexity levels / 4 domains
736d46c verified
|
Raw
History Blame Contribute Delete
2.44 kB

SCM-SQL — Dataset Statistics

Snapshot as of the 1.0 release (2026-07-30).

Totals

  • Pairs: 500
  • Turn-level trials: 556 (L6 pairs are multi-turn)
  • Complexity levels: 6 (L1 → L6)
  • Domains: 4 (demand, finance, inventory, logistics)

Per-level distribution

Level Style Pairs Turn-level trials
L1 Single-table filter / aggregate 100 100
L2 Two-table JOIN + GROUP BY 100 100
L3 Nested subquery / semi-join / 3-table join 130 130
L4 Window function 60 60
L5 CTE + rollup 60 60
L6 Multi-turn refinement (2-3 turns each) 50 106
Total 500 556

Per-domain coverage

A single pair can carry multiple domain tags — cross-domain pairs (especially L3 and L6) count in every domain they touch.

Domain Approx. pairs tagged
demand ~ 145
finance ~ 140
inventory ~ 130
logistics ~ 130

Provenance

Bucket Count How it was authored
Hand-crafted core 100 Author-written (NL + gold SQL) covering the queries a supply-chain analyst asks in practice
LLM-generated (execute-verified) 400 Generated by gpt-4o-mini from few-shots of the hand-crafted core; every candidate must (a) parse via sqlglot, (b) execute against the Odoo demo, and (c) return a non-empty non-null result set before acceptance

Reference-implementation performance on this dataset

Numbers below are for reference only — they are what the reference multi-agent NL-to-SQL pipeline at github.com/AniruddhaPKawarase/scm-nl2sql achieves on SCM-SQL. Your system will produce different numbers.

Metric Ours (gpt-4o-mini backbone) MAC-SQL comparator (same backbone)
Execution Accuracy (EX) 11.5 % 10.1 %
Soft-EX 11.9 % 10.1 %
95 % bootstrap CI (EX) [8.8 %, 14.4 %] [7.7 %, 12.8 %]
Wilcoxon signed-rank p 0.325

Interpretation: the aggregate execution accuracy is intentionally low. SCM-SQL is a hard benchmark by construction — 498-table Odoo schema, JSONB translated-name fields, multi-turn dialogues, multi-domain composition. This is a feature, not a bug: a benchmark you never fail is not testing you.