An evaluation set of 500 distinct (question, predicted SQL) items from the CoRE-DB pilot on BIRD airline and Spider driving_school / allergy_1. Each is labelled WRONG or CORRECT with executed evidence under the schemas' structural invariants. It gates reward-database builds: FP = wrong queries accepted; FN = correct queries rejected.

Counts

DB	label	tier	items
airline	CORRECT	EVIDENCE_S	43
airline	CORRECT	PROOF	29
airline	WRONG	WITNESS_ORIGINAL	35
airline	WRONG	WITNESS_S	35
allergy_1	CORRECT	EVIDENCE_S	57
allergy_1	CORRECT	PROOF	38
allergy_1	WRONG	WITNESS_ORIGINAL	50
allergy_1	WRONG	WITNESS_S	46
driving_school	CORRECT	EVIDENCE_S	50
driving_school	CORRECT	PROOF	33
driving_school	WRONG	WITNESS_ORIGINAL	42
driving_school	WRONG	WITNESS_S	42

Total: 500 items.

DB	spurious on original
airline	35
allergy_1	46
driving_school	42

Review columns

SQLs are complete. Result counts and heads refer to the WRONG witness, or to the original DB for CORRECT. Equal counts alone do not imply equal results; consult verdict_reason and evidence_json.

item_id: Distinct source item identifier; qid is also preserved in raw JSONL.
db: Schema: airline, driving_school, or allergy_1.
dataset: Source benchmark: bird or spider.
question: Complete natural-language question.
gold_sql: Complete reference SQL, unchanged.
pred_sql: Complete predicted SQL, unchanged.
label: Stored WRONG or CORRECT label.
label_tier: Stored tier: WITNESS_ORIGINAL, WITNESS_S, PROOF, or EVIDENCE_S; finite agreement is not universal proof.
spurious_on_original: WRONG and explicitly equal on the original DB; see the version caveat below.
verdict_reason: One deterministic sentence explaining the recorded verdict.
witness_env: WRONG witness environment (O = original); CORRECT uses O for displayed results.
witness_state_sha8: First eight hexadecimal characters of the displayed state hash.
gold_n_rows: Gold row count in the displayed state.
pred_n_rows: Prediction row count in the same state.
gold_head: Gold stored result head as JSON, capped at 200 characters.
pred_head: Prediction stored result head as JSON, capped at 200 characters.
sql_diff_hint: Token diff (prediction versus gold), case-folded; capped at 160 characters.
n_rollouts: Number of source rollouts represented by this item.
policies: Policy rollout counts.
question_split: Source question partition; Q_train is training-visible.
label_sources: Joined source label provenance.
invariants_version_short: Invariant description before the first semicolon, capped at 80 characters.
evidence_json: Full nested evidence JSON, unchanged in content; last column.

Labels

WRONG / WITNESS_ORIGINAL has a certified separation on the original DB; WRONG / WITNESS_S matches there but has a certified separation on an S2-valid repaired state (certificate cases a/b/c).
CORRECT / PROOF requires literal-preserving canonical identity, strict literal checking, and executed equality on the original and checked invariant-valid states.
CORRECT / EVIDENCE_S is a non-identical form with finite S2-valid-state agreement and tie checks; k includes the original DB, and member hashes identify the checked states. It is test-suite evidence, not a proof for every legal DB.
These are stored evidence labels; packaging makes no SQL executions, label changes, or LLM calls. Version-specific construction details follow in the full card.

Load

from datasets import load_dataset
items = load_dataset("thanhdath/coredb-eval500-pilot3", split="eval")

Access follows this repository's existing gating policy; authenticate an approved HF account.

Provenance

Source. The 12,800 evaluation rollouts (diag/w11/W11_REGRESSION_ROWS_v1.jsonl.gz). They come from two GRPO policies, ts_plus_r and A, with 6,400 rollouts each, on the pilot's Q_train questions: airline 46, driving_school 54, allergy_1 62.
Item. One distinct (db, question id, predicted SQL). The field rollouts lists every rollout index with that SQL.
Construction separation. The items are disjoint from the CoRE-DB build TRAIN set (B2_TRAIN_SPLIT_v1.1.json, 10879 of the 12,800 rollouts). EVAL500 never enters construction.
Question overlap. The questions are not disjoint: 149 of the 149 EVAL500 questions also have other rollouts in TRAIN.
Pre-registration. diag/w23/W23_HYPOTHESES_v1.md, amendments 3 (commit 14fb3a6f4) and 3a (commit 8ee62a176). Both were committed before the draw.

Packaging release: v1.1; previous HF source revision 740f2524; HF release tags v1.0 and v1.1 (publication results recorded in the local review report). The exact input JSONL and manifest are in raw/; rollout indices, qids, and source fields remain there. The full nested evidence is retained in parquet. The 12,800 source rollouts include policies A and ts_plus_r; these are training-visible Q_train questions, not a held-out-question test.

Limitations

EVIDENCE_S is not a proof. It is a test suite of up to 223 states. If no state varies the relevant data, an inequivalent prediction can agree on all of them. k_distinct_gold_results shows how much the states vary the gold's answer.
S is a hypothesis about legal DBs. It is mined from one original DB per schema.
- The WITNESS_S witnesses are synthetic S2-valid states.
- S does not encode real-world value mappings such as airport name <-> code. A WITNESS_S item can therefore differ only because the gold hard-codes such a constant. Example: airline q5867, where the gold filters on DEST = 'MDW' and the airport name, and the prediction filters on the name only.
Tie certificate scope. It covers the outer ORDER BY / LIMIT. Bare-column nondeterminism (SQLite picking an arbitrary row for a non-aggregated column) is not certified.
Wrong means wrong for this gold and comparator. The comparator is order-sensitive when the gold has ORDER BY, and NULL vs 0 counts as a difference.
Question overlap. The items are disjoint from TRAIN, but the questions overlap (see Provenance).
Size. The set is small: about 70 WITNESS items and 70 to 95 CORRECT items per DB.

Files: EVAL500_v1.1.jsonl (sha256 ac4f43f03e94c5a28daabd00413e907fdcc2bc952b499e8e047ed686c260fca8), EVAL500_MANIFEST_v1.1.json. Code: diag/w23/w23_eval500_v1.py (exec11, evidence11, draw11, card11).

Full technical card, copied unchanged. BIRD and Spider source data are CC BY-SA 4.0.

Downloads last month: 26

Total file size:

3.56 MB