thanhdath/coredb-eval-v2

CoRE-DB evaluation set v2 (spurious vs. correct SQL pairs)

This dataset has 1849 (gold SQL, candidate SQL) pairs. Each pair is labelled Spurious (the candidate is not equivalent to the gold under the real database's semantics; it may still pass some instance) or Not-spurious (the candidate is equivalent). No pair is labelled Unknown.

It is built to measure reward databases for text-to-SQL RL: the original database (ORI), CoRE-DB before and after merging, and a test-suite (TS-style) environment.

Labels: {"Spurious": 1117, "Not-spurious": 732}. Splits: {"train": 1163, "test": 686}. The 60 / 40 split is stratified by DB x label x CoRE-after verdict. Inside each stratum, pairs are ordered by sha256(pair_id) and the first round(0.6 n) go to train. The split is done on normalised (db, gold, candidate) keys (whitespace / ';' / case), so every copy of a key falls in one split. A DB appears in both splits, and 187 of 795 (db, question) pairs appear in both. The test split therefore measures new candidate SQL on seen schemas and questions.

There were 3 cross-source duplicates of one key, so one row was kept per key. The winner is: executed evidence > unanimous vote > majority vote > exp 13 judgment. The losing rows are recorded in dup_dropped_rows.

Split of record from v2.5 on: split_group. Experiments started after tag v2.5 must use the column split_group; split is kept unchanged for the frozen exp 17 / 18 results.

• split_group assigns every near-identical group (same DB and gold, candidate equal after alias removal and conjunct sorting) and every normalised-SQL duplicate to ONE split. The two keys are merged by union-find into split_group_id, so no group spans both splits.
• Target 60 / 40, stratified by DB x group label x CoRE-after verdict.
• The result is 1,453 groups: train 917 groups / 1,185 rows, test 536 groups / 664 rows. 806 rows sit in a different split than in split.
• The data/train.parquet / data/test.parquet files still follow split. Use the split_group column to re-split.

v2.6. The airline "Miami" near-identical group (7 pairs; gold filters DEST = 'MIA', candidates filter Airports.Description = 'Miami, FL: Miami International') was re-voted with the fact measured on the original: "Airports.Description is NOT unique: 43 descriptions occur more than once among 6,510 rows; the gold is absolute".

• 24 of 35 votes said Spurious, so all 7 pairs are now Spurious (they were Not-spurious).
• split and split_group are unchanged.

v2.7. All bike_share_1 (136) and legislator (142) pairs, train and test, were re-judged by the same 5 models (forced binary) with verified facts stated as database facts. One label per near-identical group: the majority over all votes of the group.

• bike_share_1 facts were measured on the original: station renames for ids 25/46/47/80 (name join and id join are not equivalent), 0 dangling station ids in 658,901 trips, duration >= 60 s, 0 trips before installation, weather (date, zip_code) and station.name unique, and empty-set semantics.
• legislator facts: the exp 17 kept semantic rules with their support, plus the exp 14 inventory with the original's exceptions, restricted to the tables and columns the queries read.
• 51 labels changed:
  • bike_share_1: 31 Not-spurious -> Spurious, 10 Spurious -> Not-spurious.
  • legislator: 3 Not-spurious -> Spurious, 7 Spurious -> Not-spurious.
• 8 rows in groups with fewer than 3 votes keep their label (flag in bikeleg_rejudge): the OpenRouter credit briefly ran out, then the $6 cap was reached at $6.35.
• split and split_group are unchanged.

v2.8 (label version of record; mechanical, no LLM calls). Rule of record: a Spurious label needs a witness state that is admissible AND keeps the gold informative (non-empty, not all-NULL). This matches the reward tree's G1 gold rule.

• Pairs that differ from the gold only when the gold's row set is empty (the gold returns NULL) are Not-spurious, with v2_8_reason = "gold-uninformative-only difference" and the previous label in label_before_v2_8.
  • The airline SUM(CASE WHEN p THEN 1 ELSE 0 END) vs COUNT(..) WHERE p class: 83 pairs, all Spurious before. 76 are now Not-spurious.
  • v2.9 correction (reviewer): 7 of the 83 (44457161, 4511eeeb, 50accd88, 63fe4fda in test; 536aaa57, 77519fd1, c3ae7197 in train) are separated on the ORIGINAL DB with an informative gold (e.g. 731 vs 0). They return to Spurious with reason "refused on the original with informative gold" and label_method executed_witness_original_db. The previous label is in label_before_v2_9, and the gold's row count on the original is in v2_9_gold_n_rows_original. A generalised check executed gold and candidate on the original for all 84 relabelled pairs: 77 agree there with an informative gold, 0 have an uninformative gold, and exactly these 7 are separated.
  • bike_share_1 pair 9aef4dec66638b0c, which differs only if the station table is empty. It was already Not-spurious in v2.7.
• This supersedes the v2.3 SUM(CASE) vote outcome. split and split_group are unchanged.
• Legislator fact note (reviewer): the rule current.maplight_id -> historical.icpsr_id has 22 orphan rows on the original. It is an inferred inclusion with exceptions, not a foreign key. It was missing from the v2.7 fact packets (0 packets cited it).

Orientation of the confusion table

Positive means the environment accepts (pays) the SQL.

cell	meaning
TP	correct SQL accepted
FP	spurious SQL accepted (spurious correctness)
FN	correct SQL refused (false refusal, an instance artefact)
TN	spurious SQL refused

• precision = TP / (TP + FP): how much of what the environment pays is really correct.
• false-refusal rate = FN / (TP + FN).

Sources and label provenance (column source, label_method, label_tier)

source                    label       
eval500_v1.1              Not-spurious    276
                          Spurious        224
exp13_bird                Not-spurious     40
                          Spurious         79
exp13_spider              Not-spurious      9
                          Spurious         50
ori12_extra_bike_share_1  Not-spurious     25
                          Spurious        105
ori12_extra_legislator    Not-spurious     41
                          Spurious         89
ori12_rollouts_bird       Not-spurious    131
                          Spurious        169
ori12_rollouts_spider     Not-spurious    118
                          Spurious        181
w21_ori_accept_ts_refuse  Not-spurious     92
                          Spurious        220

• eval500_v1.1: thanhdath/coredb-eval500-pilot3 at tag v1.1. The pilot schemas are airline (BIRD) and driving_school / allergy_1 (Spider).

  • Labels come from executed evidence (EVAL500 tiers PROOF / EVIDENCE_S / WITNESS_ORIGINAL / WITNESS_S).
  • 89 pairs were re-checked against invariants mined from the original DBs. Witnesses that violate a mined rule on a queried column were sent to the LLM vote: label_before_recheck, recheck_status.

• exp13_bird / exp13_spider: 180 CoRE-only refusals of ORI-paid rollouts (ORI12 run), judged and reconciled in exp 13.

  • Class A maps to Spurious; class C maps to Not-spurious (CoRE wrongly refuses correct SQL).
  • They are selected pairs and carry no rates.

• ori12_rollouts_bird / _spider: a stratified sample of ORI12's ORI-paid rollouts (cells CoRE-accept / CoRE-refuse, one rollout per prompt). Canonical-identical SQL is labelled Not-spurious; the rest come from the LLM vote.

• w21_ori_accept_ts_refuse: pilot rollouts that the original DB pays and the TS+R suite refuses. Their evidence label is in evidence_label (refusing instances that violate an invariant mined from the original). The final label comes from the LLM vote.

• ori12_extra_bike_share_1 / ori12_extra_legislator (v2.2): 130 pairs per DB from ORI12 rollouts (steps 0-2999).

  • Distinct (prompt, SQL), one rollout per prompt per round. Stratified over ORI accept/refuse x CoRE-pinned accept/refuse with quotas of 30 per cell; deficits move to the ORI-paid cells, then ORI-refuse/CoRE-refuse (label_tier = ORI12_EXTRA_).
  • Labels: canonical identity -> Not-spurious; ORI refuses with a usable gold -> Spurious (the original DB is the executed witness); the rest from the same 5-model vote (147 pairs: 78 unanimous, 67 majority, 2 ties broken by Opus).
  • Rows published in v2.1 keep their split and label. New rows take the same key-grouped stratified 60/40 split.

• Constraint-aware re-judge (v2.3). The same 5 models gave forced binary votes with verified database facts stated in the packet (columns constraint_shown, label_before_constraint_rejudge, constraint_rejudge_flag).

  • The 22 exp 17 rule-conflict pairs, each with its verified rules (e.g. "(StuID, Allergy) is unique in Has_Allergy: 0 duplicates in 59 rows", "customer_status_code in {Good Customer, Bad Customer}"). 11 changed Spurious -> Not-spurious (allergy_1 6, driving_school 4, airline 1); 11 stayed Spurious.
  • The whole airline class SUM(CASE WHEN p ..) vs COUNT(..) WHERE p (83 pairs), with the fact "the filtered set can be empty; SUM over zero rows returns NULL while COUNT(*) returns 0; the gold is absolute". All 83 are now Spurious (82 unanimous): 52 changed Not-spurious -> Spurious, one consistent label for the class.
  • Splits are unchanged.
  • Note: for the SUM(CASE) class, the fact text itself states the outcome ("the gold is absolute"; NULL vs 0 differ), so this vote applies a stated semantic rule rather than discovering one.

• v2.4 (reviewer fixes).

  • The 4 allergy_1 pairs that join Student were re-voted with the verified fact "Has_Allergy.StuID can be absent from Student (1 orphan row: 1013, 'Ragweed')". All 4 changed back to Spurious (unanimous).
  • The 4 driving_school flips carry label_flag low_support_rule: the enum is verified on 15 rows.
  • Near-identical pairs share one label. Pairs are grouped by the same DB, gold, and candidate after removing aliases and sorting conjuncts (column near_key). In a mixed group an executed-evidence label wins, else the majority, else the label with more unanimous votes, else Spurious. This touched 18 groups and 21 rows (old label in label_before_group).
  • Splits stay as published. 104 near-identical groups span both splits, so near-duplicates of a test pair can be in train.

LLM vote (OpenRouter)

• Five models, one forced binary answer each (Spurious / Not-spurious, with a one-line reason), majority decides.
• judge_flag is unanimous / majority / tie_opus. A tie arises only from a missing vote. It is broken in the order Opus > GPT > Gemini > DeepSeek > Sonnet; none occurred in the final run. 11 responses were not JSON, and their label was taken from the label word in the text.
• Settings: temperature 0, max_tokens 1500, timeout 60 s, at most 2 retries.
• Packets are blind: question, evidence, schema of the queried tables, gold, candidate, and both results on the original DB. They carry no environment verdict and no instance names.
• Rubric: the gold is absolutely correct. A candidate is Not-spurious iff it returns the gold's result on every legitimate instance of the schema.

{
 "returned_models": [
  "anthropic/claude-opus-5.5",
  "anthropic/claude-sonnet-5.5",
  "deepseek/deepseek-v4-pro",
  "google/gemini-3.8-flash",
  "openai/gpt-5.6-luna"
 ],
 "flags": {
  "unanimous": 665,
  "majority": 258
 },
 "spend_usd": 0.2361,
 "pairwise_agreement": {
  "anthropic/claude-opus-5.5 | anthropic/claude-sonnet-5.5": {
   "n": 923,
   "agree": 805
  },
  "anthropic/claude-opus-5.5 | openai/gpt-5.6-luna": {
   "n": 923,
   "agree": 848
  },
  "anthropic/claude-opus-5.5 | deepseek/deepseek-v4-pro": {
   "n": 921,
   "agree": 812
  },
  "anthropic/claude-opus-5.5 | google/gemini-3.8-flash": {
   "n": 923,
   "agree": 743
  },
  "anthropic/claude-sonnet-5.5 | openai/gpt-5.6-luna": {
   "n": 923,
   "agree": 798
  },
  "anthropic/claude-sonnet-5.5 | deepseek/deepseek-v4-pro": {
   "n": 921,
   "agree": 788
  },
  "anthropic/claude-sonnet-5.5 | google/gemini-3.8-flash": {
   "n": 923,
   "agree": 829
  },
  "openai/gpt-5.6-luna | deepseek/deepseek-v4-pro": {
   "n": 921,
   "agree": 795
  },
  "openai/gpt-5.6-luna | google/gemini-3.8-flash": {
   "n": 923,
   "agree": 738
  },
  "deepseek/deepseek-v4-pro | google/gemini-3.8-flash": {
   "n": 921,
   "agree": 755
  }
 }
}

Environment verdict columns

All verdicts use the trainer comparator: a set of rows, tagged per member, where every member must agree.

column	environment
ori_verdict	ORI
core_before_verdict	CoRE before merging: the full v7 candidate pool
core_pinned_verdict	CoRE after merging: tree 69a2e609
ts_verdict	pilot TS+R suite, set comparator; n/a outside the pilot DBs
ts_bag_verdict	the pilot trainer's bag comparator

Headline table

By source

group	env	TP (correct paid)	FP (spurious paid)	FN (correct refused)	TN (spurious refused)	n/a	precision	false-refusal rate	spurious-acceptance rate
eval500_v1.1	ORI	276	107	0	117	0	0.721	0.000	0.478
eval500_v1.1	CoRE before merging (v7 pool, all units)	268	44	8	180	0	0.859	0.029	0.196
eval500_v1.1	CoRE after merging (pinned 69a2e609)	273	64	3	160	0	0.810	0.011	0.286
eval500_v1.1	TS-style (set comparator)	244	25	32	199	0	0.907	0.116	0.112
eval500_v1.1	TS-style pilot trainer (bag comparator)	244	7	32	217	0	0.972	0.116	0.031
exp13_bird	ORI	40	79	0	0	0	0.336	0.000	1.000
exp13_bird	CoRE before merging (v7 pool, all units)	0	0	40	77	2		1.000	0.000
exp13_bird	CoRE after merging (pinned 69a2e609)	0	0	40	79	0		1.000	0.000
exp13_bird	TS-style (set comparator)	0	0	0	0	119
exp13_bird	TS-style pilot trainer (bag comparator)	0	0	0	0	119
exp13_spider	ORI	9	50	0	0	0	0.153	0.000	1.000
exp13_spider	CoRE before merging (v7 pool, all units)	0	0	9	48	2		1.000	0.000
exp13_spider	CoRE after merging (pinned 69a2e609)	0	0	9	50	0		1.000	0.000
exp13_spider	TS-style (set comparator)	0	0	0	0	59
exp13_spider	TS-style pilot trainer (bag comparator)	0	0	0	0	59
ori12_extra_bike_share_1	ORI	25	61	0	44	0	0.291	0.000	0.581
ori12_extra_bike_share_1	CoRE before merging (v7 pool, all units)	19	29	6	76	0	0.396	0.240	0.276
ori12_extra_bike_share_1	CoRE after merging (pinned 69a2e609)	20	31	5	74	0	0.392	0.200	0.295
ori12_extra_bike_share_1	TS-style (set comparator)	0	0	0	0	130
ori12_extra_bike_share_1	TS-style pilot trainer (bag comparator)	0	0	0	0	130
ori12_extra_legislator	ORI	41	37	0	52	0	0.526	0.000	0.416
ori12_extra_legislator	CoRE before merging (v7 pool, all units)	37	13	4	76	0	0.740	0.098	0.146
ori12_extra_legislator	CoRE after merging (pinned 69a2e609)	37	16	4	73	0	0.698	0.098	0.180
ori12_extra_legislator	TS-style (set comparator)	0	0	0	0	130
ori12_extra_legislator	TS-style pilot trainer (bag comparator)	0	0	0	0	130
ori12_rollouts_bird	ORI	131	169	0	0	0	0.437	0.000	1.000
ori12_rollouts_bird	CoRE before merging (v7 pool, all units)	107	36	23	132	2	0.748	0.177	0.214
ori12_rollouts_bird	CoRE after merging (pinned 69a2e609)	109	41	22	128	0	0.727	0.168	0.243
ori12_rollouts_bird	TS-style (set comparator)	0	0	0	0	300
ori12_rollouts_bird	TS-style pilot trainer (bag comparator)	0	0	0	0	300
ori12_rollouts_spider	ORI	118	181	0	0	0	0.395	0.000	1.000
ori12_rollouts_spider	CoRE before merging (v7 pool, all units)	106	26	3	150	14	0.803	0.028	0.148
ori12_rollouts_spider	CoRE after merging (pinned 69a2e609)	115	34	3	147	0	0.772	0.025	0.188
ori12_rollouts_spider	TS-style (set comparator)	0	0	0	0	299
ori12_rollouts_spider	TS-style pilot trainer (bag comparator)	0	0	0	0	299
w21_ori_accept_ts_refuse	ORI	92	220	0	0	0	0.295	0.000	1.000
w21_ori_accept_ts_refuse	CoRE before merging (v7 pool, all units)	72	82	20	138	0	0.468	0.217	0.373
w21_ori_accept_ts_refuse	CoRE after merging (pinned 69a2e609)	82	128	10	92	0	0.390	0.109	0.582
w21_ori_accept_ts_refuse	TS-style (set comparator)	0	17	92	203	0	0.000	1.000	0.077
w21_ori_accept_ts_refuse	TS-style pilot trainer (bag comparator)	0	0	92	220	0		1.000	0.000

Caveats:

• The exp 13 pairs are selected by ORI-accept / CoRE-refuse, and the w21 pairs by ORI-accept / TS-refuse. Their CoRE or TS cells are fixed by construction.
• 23 EVAL500 labels changed from WRONG/WITNESS_S (Spurious) to Not-spurious: 22 airline, 1 allergy_1. Their executed witness violates a rule mined from the original on a queried column, and the 5-model vote then overrode the witness. EVAL500 is therefore 273 Not-spurious / 227 Spurious here, not 250 / 250. The old label is kept in label_before_recheck.
  • Consequence: on the EVAL500 pairs the TS-style FN goes from 6 to 29. The jump is mainly these flips: the TS suite refuses them because they are separated there. It is not a new TS weakness.
• Of the 312 w21 pairs, 69 are Not-spurious by the vote. The rule-break heuristic had said Not-spurious for 282; 67 of those survive, plus 2 that the heuristic left open.
• The phase-2 ORI12 pairs are all ORI-paid, so ORI has no refusal cells there.
• EVAL500's EVIDENCE_S tier is itself a TS-like criterion.
• The rate rows over the pilot use 3 schemas.