README.md

---
license: cc-by-4.0
task_categories:
- text2text-generation
tags:
- text-to-sql
- reinforcement-learning
- database
---

# CoRE-DB SQL-R1

A compact, gold-refined, counterexample-enriched reward database for reinforcement-learning
text-to-SQL training. It is built from SQL-R1's own SynSQL-complex-5K training corpus, over the
benchmark's own gold queries only — no model-written query is an input to construction. The
database is meant to sit under an execution-based RL reward: SQL is judged by running it, and
this release is the tree that reward runs against.

**Access is gated with manual approval.** Request access on the repo page; approvals are granted
by the repo owner.

## What is in the tree

The build follows three steps, run only from the gold SQL:

1. **Compact and populate.** A fresh copy of the read-only control (SQL-R1's own SynSQL
   databases) populated, database by database, until every gold of all 5,000 corpus rows returns
   rows under both gold forms it is evaluated in.
2. **Random instances per database**, drawn once and shared by that database's golds — schema-valid,
   FK-closed, multi-row, with NULLs, duplicates, groups, ties, and boundaries around the gold's own
   literals.
3. **The gold and its neighbours** (single-edit query mutations) executed on those instances; where
   a neighbour ties the gold on the compact tree but a random instance separates them, the rows that
   carry the difference are taken from that instance, minimized, closed over their own foreign keys,
   and planted back into the compact tree, parents first, with a re-check after every database.

No gold SQL was ever modified in this build (step 2 — gold hardening — was not applied to this
arm, by decision). The held-out policy decodes used to measure catch elsewhere are never an input
to construction.

## Identity

| | |
|---|---|
| content fingerprint | `b22549ce81c383b3c1531f2254ea3159` |
| archive | `release_b22549ce.tar.zst` |
| archive sha256 | `0f3d4b848c03dfc1acbe9d69db2497a3d4299177a6c0fa26fb2bf27285ff4ae6` |
| databases | 4,064 |
| total tree size | 221,417,472 bytes (221.42 MB) |
| rows (compact tree) | 88,740 |

## Key statistics (from the freeze bundle)

**Corpus.** 5,000 gold queries (4,000 train / 1,000 test split of the vendored SynSQL-complex-5K),
over 4,064 databases.

**Step 1 — constructive population.**

| | |
|---:|---:|
| golds total | 5,000 |
| golds non-empty | 4,998 |
| golds empty | 2 (proved unanswerable on any faithful instance — kept in the corpus, never subsampled) |
| golds unexecutable | 0 |
| rows before | 75,913 |
| rows after | 88,740 |
| rows added | 12,827 |
| databases malformed / FK-orphans-increased / new declared violations | 0 / 0 / 0 |

**Step 3 — separation and planting.**

| | |
|---:|---:|
| databases swept | 3,065 / 4,063 (75.4%) |
| variants tied on the compact database | 19,333 |
| variants separated by some random instance | 8,352 (43.2%) |
| variants no instance told apart | 10,981 (56.8%) |
| pairs planted (transplanted) | 322 |
| rows the tree gained over step 1 | +12,827 |

A database the sweep never reached is unresolved, not refused: its golds are still populated by
step 1 and still gated; what it lacks is step-3 enrichment.

**Gate, over all 5,000 golds, in both gold forms (published and the reward's executed read
path):** 4,998 / 5,000 non-empty, 0 malformed databases, 0 FK-orphan increases, 0 new declared
violations. GATE CLEAN.

The control (SQL-R1's own SynSQL databases, unmodified, used by the matched-pair comparison) has
content fingerprint `a8482f003ec7fdd5ebbfbd2842231f0a`, 4,064 databases, 221,409,280 bytes.

## Restore and verify

```bash
# 1. verify the archive against the sha256 recorded above
sha256sum -c <(echo "0f3d4b848c03dfc1acbe9d69db2497a3d4299177a6c0fa26fb2bf27285ff4ae6  release_b22549ce.tar.zst")

# 2. restore
zstd -d release_b22549ce.tar.zst -c | tar -x -C <dest>

# 3. recompute the content fingerprint (script vendored in the source repo at
#    campaigns/chained_05b_tkde/mining_v2/v7/coredb/fingerprint.py)
python3 fingerprint.py --bird_root <dest>/tree --spider_root <dest>/empty_dir --label coredb_sqlr1
# expect: content_fingerprint == b22549ce81c383b3c1531f2254ea3159, n_databases == 4064
```

Note: the tree mounts under the `bird_root` argument only — this dataset routes through the
same bird|spider|spider2 mount scheme as the other CoRE-DB release, and SynSQL is labeled `bird`
as a routing label only (it does not affect content).

## Files in this repo

- `release_b22549ce.tar.zst` — the frozen tree archive (4,064 SQLite databases) plus the step-2
  gold-format parquet.
- `FREEZE_BUNDLE.md` — the full freeze record with every measured number and how it was produced.
- `RELEASE_SQLR1.json` — the machine-readable release record.