thanhdath/bird-spider-rl-hard-train

BIRD and Spider training samples selected by model difficulty

The training subset contains 9,687 examples: 5,100 BIRD train and 4,587 Spider train. It removes examples answered correctly in all eight observed samples by the released FINER-SQL 0.5B BIRD model. Difficulty is measured from model outputs, rather than SQL length or a required number of joins.

This is a dataset release requested by its owner. The selection rule has not been established to preserve or improve downstream accuracy; the earlier small pilot described below failed its Spider accuracy criterion.

Configurations

Configuration	Rows	BIRD train	Spider train
train_hard (default)	9,687	5,100	4,587
train_all_labelled	18,087	9,428	8,659
train_hard_ori_screened	9,687	5,100	4,587
train_hard_core_screened	9,687	5,100	4,587
spurious500_ori	500	329	171
spurious500_core	500	329	171

The default configuration is the selected training set. The second configuration preserves every source example, including those excluded from training, so the selection can be inspected. Neither configuration contains a dev/test split.

from datasets import load_dataset

train = load_dataset("thanhdath/bird-spider-rl-hard-train",
                     "train_hard", split="train")

Selection rule

k_init counts original-database execution matches among eight samples from griffith-bigdata/FINER-SQL-0.5B-BIRD. k_ref is the same measurement for a 0.5B policy continued on original databases, AP24 checkpoint 828.

Keep an example exactly when:

gold_valid and k_init <= 7 and (k_init >= 1 or k_ref >= 1)

• 8,802 examples have 1 <= k_init <= 7: the initialization sometimes succeeds and sometimes fails.
• 885 examples have k_init = 0 and k_ref >= 1: the initialization failed all eight draws, while the reference policy solved at least one draw.
• The selected set contains zero examples with k_init = 8.

All original source fields and selected row order are preserved: dataset_name, split, prompt_raw, messages, question, sample_id, db_id, and groundtruth_sqls. The subset adds k_init, k_ref, and bin. It includes the full prompt messages, question and gold SQL; it does not bundle database files.

The full labelled configuration additionally contains the membership flag r2_keep, gold execution status and exclusion reasons, and SQL-structure descriptors. Structural descriptors are for inspection; they do not filter rows.

Exclusions and interpretation

Bin in the full source corpus	Rows	Included in train_hard
LEARNABLE: valid gold, 1–7 initialization matches	8,802	Yes
HARD_reachable: valid gold, 0 initialization matches, reference succeeds	885	Yes
EASY: valid gold, 8 initialization matches	2,528	No
NEVER_SOLVED: valid gold, no match in either policy's eight draws	3,265	No
INVALID: unusable gold execution under the fixed rules	2,607	No

Eight successful draws do not prove an example will always be correct. NEVER_SOLVED means zero of sixteen observed draws, not that learning the example is impossible. Difficulty is specific to the two measured 0.5B policies; it has not been measured for 1.5B or 3B models.

Gold validity excludes execution errors/timeouts at the reward limit, empty results, all-NULL results, and results reaching the 10,000-row cap. The full configuration preserves all 366 rows with unknown k_init/k_ref values as null, rather than replacing unknown values with zero. Existing ORDER/LIMIT tie and bare-aggregate flags are retained for inspection and are not additional selection filters.

Execution agreement is measured on original training databases, not CoRE databases. A match is not a proof of semantic equivalence. The frozen labeller uses BIRD set-result comparison and Spider's evaluation-aligned comparison, including its existing DISTINCT handling. This release changes neither SQL nor the comparator and introduces no DISTINCT-specific selection rule. Selecting this dataset is not evidence that CoRE training outperforms original-DB training.

Provenance and reproducibility

The source is an existing joint BIRD/Spider training corpus of 18,087 rows. The initialization labels use the released FINER-SQL weights, not the earlier SQL-Writer SFT model. Both policies were decoded with eight samples per prompt, temperature 1.0, top-p 1.0, seed 4242, and maximum completion length 2,048 tokens: 144,696 recorded samples per policy. No new model generation, sampling sweep or training run was performed for this upload.

• Initialization weights MD5: 22c730ffe79fdcdd7a935ff6f03d790e.
• Reference weights SHA-256: 38b89afd56b61410e0b3ba72590f24242f7c25b74fd56955befc46ec78be41f1.
• Source Arrow SHA-256: 2e203ca0789f7b6f23c466bd23d9cd1286e437586defb1345736139060c1db16.

STATS_v1.json records file hashes and composition; VALIDATION.json records independent membership, label-provenance, train-split and exact source-value checks. Scalar Arrow strings widen to Parquet large_string; logical source values and nested list fields are preserved.

Reference-screened configurations (2026-10-23)

train_hard_ori_screened and train_hard_core_screened are the same 9,687 rows, in the same order and with the same columns as train_hard. Only groundtruth_sqls differs. Its first entry is still the original BIRD/Spider gold, unchanged. The entries after it are the alternative reference SQLs that the atomic-operations reward compares against.

How the screen works. Each reference was run as if it were a model prediction, using the FINER-SQL trainer's own verdict function (check_sql_correct in grpo_writer.py: 60 s timeout, 10,000-row cap, row-set comparison, and the rule that an unusable gold can never certify a prediction). Each reference was compared against the gold on that configuration's databases. A reference was kept only when the trainer judged it correct. References that returned different rows, failed to run, or timed out were removed.

• train_hard_ori_screened: screened on the original BIRD-train and Spider-train databases (215 databases).
• train_hard_core_screened: screened on the CoRE bundle tree BJ (724 SQLite databases, digest 69a2e60923e2d900764fda6d9c88265c; 215 base databases, each bundled with its CoRE members). A reference is kept only when its rows equal the gold's on the base database and on every bundle member.

ORI arms must train on train_hard_ori_screened, and CoRE arms on train_hard_core_screened. The two reference lists differ because a reference can match the gold on the original data but not on the CoRE databases.

	ORI screened	CoRE screened
Rows	9,687	9,687
References in	122,714	122,714
References kept	122,710	106,808
References dropped (result mismatch)	4 (0.003%)	15,906 (12.962%)
Dropped: execution error / timeout / unusable gold	0 / 0 / 0	0 / 0 / 0
BIRD references kept / in	78,169 / 78,173	67,273 / 78,173
Spider references kept / in	44,541 / 44,541	39,535 / 44,541
Rows left with the gold only (screen)	0	261
Rows with the gold only already in train_hard	468	468

Checks: every one of the 8,055 distinct golds is judged correct when scored as its own prediction in both settings, and 50 randomly chosen kept references gave the same verdict when scored a second time. All 4 ORI drops are references for one BIRD question (sample 8323, mondial_geo). Its gold selects two columns named Name; these merge into one column in the comparator's row records, so no two-column reference can match it.

File	SHA-256
data/train_hard_ori_screened.parquet	a787785cac513f08f22596fb3e49eb14006631f0761133f013e278333e656d04
data/train_hard_core_screened.parquet	3d005e30ad9563c8269551a4c96d222438a8a8f6f953f5156b851184b0b29c31
screen receipt, ORI (RECEIPT_ori.json)	53bdc9b54b4f606d74219eb0a775adf7c5d142ea423f6d2ecca312b558e60c13
screen receipt, CoRE (RECEIPT_core.json)	7fb01d5f552b994987b9107329039b79eb5f39fe268920ef9721f5ef646671b6
input data/train_hard.parquet	1e5a7dd375424c3dd21a742ffed4a0480393b514fa87f9004ec52a80b21b7fe7

Prior pilot: accuracy was not preserved on Spider

An earlier, different experiment applied this fixed rule to a 2,208-example random sample. It retained 1,419 examples; this 9,687-example full subset has not itself been evaluated in a matched training experiment.

Original-DB arm	Training steps	BIRD dev / 1,534	Spider dev / 1,034
Full random sample, APRAND	1,104	602 (39.24%)	694 (67.12%)
Selected random subset, AHRAND	712	613 (39.96%)	674 (65.18%)

The selected pilot used 35.5% fewer optimizer steps, gained 0.72 percentage points on BIRD, and lost 1.93 points on Spider. It failed the fixed criterion of at most 0.5 points lost on either benchmark. Step reduction alone is not an end-to-end speedup or accuracy improvement. Selection costs are separate from training cost.

The source questions and SQL retain their upstream BIRD and Spider terms; this release does not assign a new license to those source contents.

Spurious-correctness subsets (spurious500_ori, spurious500_core)

These are 500 prompts from train_hard, picked for having the most references that match the gold on the original databases but not on the CoRE databases. For each prompt, the selection counts that prompt's references in the CoRE screen's dropped list. Prompts are ranked by that count, then by the dropped fraction, then by sample_id (ascending), then by dataset_name (ascending). Prompts whose gold is not usable in either GT cache, and prompts with no dropped reference, are excluded. Both configurations hold the same 500 rows in the original file order. spurious500_ori takes each row's references from train_hard_ori_screened, and spurious500_core takes them from train_hard_core_screened. Dropped references per prompt: min 11, median 16, max 38. The rule, counts and examples are in selection/SELECTION.md, and the hashes are in selection/RECEIPT.json.