Skip to content

datumwise/ground-truth-benchmark

Repository files navigation

Ground Truth benchmark — v1 publication kit (FROZEN)

This is the complete, frozen v1 kit behind the published study:

Silent Failures in Single-Shot Text-to-SQL: A Blind, Nine-Model Cross-Comparison on the "Ground Truth" Benchmark — Wang, Huayin (2026). DOI: https://doi.org/10.5281/zenodo.21349581

Layout

  • benchmark/ — the benchmark itself, as-frozen: generator (v0.3.0-g4), harness/grader, the CI reference instance dc81f6e0c4182384 (300K; warehouse, latent ground truth, 111-question manifest with answer keys, schema docs), release commitments, run guides.
  • study/ — the nine-model study kit: the exact model-facing inputs (INSTRUCTIONS.md, questions.md, schema_reference.md), all nine raw answer sets (verbatim), the scoring drivers, the committed comparison artifacts, and FINDINGS.md.
  • tables/comparison.csv — the published comparison tables as data.
  • reproduction/REPRODUCTION_LOG.md — an independent re-scoring of all nine answer sets through the harness's own path, matching the committed tables exactly (2026-07-14).
  • ERRATA.md — the grading/key history, stated plainly.
  • MANIFEST-CHECK.md / GAPS.md — mechanical inventory against the publication manifest; every item is a path or an explicit gap with a reason.

Reproduce the published numbers (one command per model)

cd benchmark
pip install duckdb pandas
python score_uploaded.py instances/dc81f6e0c4182384 ../study/results/answers_<model>.json --label <model>

Expected: the trust scores and silent counts in tables/comparison.csv, exactly.

Freeze & contamination policy

v1 is published here in full — questions, ground truth, answer keys, raw model answers — for reproducibility and audit. It is therefore retired as an evaluation instrument: once a benchmark's answers are public, future models may absorb them, and scores on it stop meaning what they meant. Future scored runs use a private revision of the benchmark, for the obvious reason. (The sealed competition holdout referenced in benchmark/PICKUP.md remains sealed and is not part of this kit — by design.)

A note on names

The engine under study was renamed after this benchmark was frozen; artifacts herein use the historical name ("Coframe" / coframe_def.cf, coframe-benchmark). They are published as-frozen: renaming inside a frozen research artifact would falsify the record. The current project is Columna, by datumwise (https://datumwise.ai).

Licenses

Documents and data: CC-BY-4.0. Code (generator, harness, drivers): Apache-2.0. See LICENSE-NOTICE.md; full license texts accompany the public repository.

About

Ground Truth benchmark v1 — the frozen publication kit behind the nine-model single-shot text-to-SQL study (DOI 10.5281/zenodo.21349581). CC-BY-4.0 docs/data, Apache-2.0 code.

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages