Hugging Face's logo

Datasets:
thanhdath
/
coredb-v7-eb380f

Tags:
text-to-sql
sqlite
reward-databases
License:
Dataset card Files
xet
 Community

You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

CoRE-DB v7 (manifest eb380ff99c52)

Frozen 2026-10-21 15:48:59. Reward-database bundles for the 1,881 BIRD + Spider training golds used by the CoRE-DB RL reward (v12 bundle adapter). Backup of the frozen tree; no model checkpoints are stored here. Built without any LLM call; every query ran through the project's db_execution API.

Contents

  • tree/{bird,spider}/<name>/<name>.sqlite: 679 SQLite files = 199 original DBs (D0) + 480 members. Member sha256 checked against inputs/BUNDLES.json member_sha256 before upload (679 named files, 0 mismatches).
  • inputs/BUNDLES.json (db_members, member_sha256, gold_index; sha256 cfd944bf0333a986), inputs/MASKS_v3.json (all empty; sha256 43d7f67ddbace000).
  • MANIFEST_v1.json (per DB, sha256 eb380ff99c52cb35741adf8ffca36f60933b560ff4c73fff734e0faf513da975), COREDB_v7_v1.json (totals), COREDB_v7_v1.md (build note).
  • scripts/: the exact build scripts (v7_prep_v1.py tasks, v7_worker_v1.py per-shard re-certification, v7_select_freeze_v1.py selection + freeze, v7_ict_job_v1.pbs cross-engine job, hf_backup_v7_v1.py this upload).
  • SHA256SUMS: sha256 of every file in this repo except README.md and SHA256SUMS itself.

Method

  • Candidate units: the retained pre-merge v4 pool, 8,002 units over 68 BIRD and 131 Spider DBs (the pool v6 selected from).
  • Near-miss pairs: 42,695 unique (gold, prediction) pairs = every construction rollout of the golds' questions
    • every CP22 rollout that is original-correct under the official judge.
  • Verdict: the official one (BIRD set(tuples); Spider ts_eval remove_distinct + result_eq, order only if the gold has ORDER BY). Paid = official-correct on D0. Separable = official-wrong on a unit where the gold executes, returns < 10,000 rows, is non-empty and not all-NULL.
  • Selection: v6 greedy, K <= 20, ties to fewer added rows, D0 first; units that official-reject a v4 PROOF/EVIDENCE guard are ineligible (104 units in 10 DBs).
  • Pairs: paid on D0 26,837; official-separable 10,933 (BIRD 4,512, Spider 6,421), all covered by v7; official-inseparable 13,308 (279 DISTINCT-only).

Quality: v6 vs v7 (same official executions)

v6 v7
members (excluding D0) 670 (BIRD 384, Spider 286) 480 (BIRD 285, Spider 195)
golds with >= 1 official-separating active member 1049 (55.8 %) 1149 (61.1 %)
BIRD / Spider 47.7 % / 69.8 % 53.9 % / 73.5 %
golds that gain a separating member v6 lacked 100 (BIRD 74, Spider 26); lost 0
members kept / dropped / new vs v6 195 / 475 / 285

Caveats (do not over-claim)

  • Separable does not mean wrong: 83 % of the separated paid pairs (9,428 of 11,353) are CP22 rollouts that are correct on the original DB; a member that separates them refuses those predictions.
  • No latency trim (v6 trimmed to 4.0x D0); v7 has at most 13 members per DB, latency measured separately.
  • One sqlite engine per DB at freeze: gf-6000 3.45.3 or gf-henry 3.50.2. The ICT cross-engine check is pending.
  • 18 DBs have D0 only (no separable pair); training uses the single-DB reward on them.
  • Masks: 0 golds masked (76,959 / 76,959 gold-unit validity checks pass).
Downloads last month
10
Total file size:
212 MB