CoRE-DB value-filling agents
Backup of the multi-agent per-database value-filling artefacts from
campaigns/chained_05b_tkde/valuefill_agents/ in thanhdath/finer-sql-tkde (private repo).
This is the offline-proxy measurement harness and its outputs backing the paper's agentic
value-filling section, not a training corpus by itself.
What this answers
The CoRE-DB construction fills each compact database with random instances to separate a
gold SQL query from its near-miss neighbours (edits that would otherwise collect reward
spuriously). One shared random-value generator (v18/fuzz_lane/generate.py) fills every
database today. This package asks whether a per-database value-filling program —
written by a Coder agent and accepted only after a Verifier agent proves it by execution —
separates more (gold, near-miss) pairs than the shared generator, on the same pairs, same
instance budget, and never separates fewer.
Run of record
2026-09-15 .. 2026-09-17, model claude-sonnet-5. Full sweep over both corpora's eligible
train databases:
| corpus | eligible | swept | verified module | unverified |
|---|---|---|---|---|
| BIRD train | 72 | 52 | 42 | 10 |
| Spider train | 167 | 116 | 106 | 10 |
| total | 239 | 168 | 148 | 20 |
An unverified database has no fill_rules.py; it falls back to the shared generator. Its
meta.json, verifier_report.json and attempts/ are kept — the repair rounds that
failed are as much the audit trail as the ones that succeeded.
Full-sweep headline (held-out pairs, offline-proxy label)
Held-out = pairs not used to construct/verify the evaluated database's module (the construction-ceiling numbers, run on the in-sample pairs, are in the same eval files).
| corpus | pairs | shared catch | agent catch | delta |
|---|---|---|---|---|
BIRD (out/full_bird/eval_heldout.json) |
1,607 | 29.68 % | 31.86 % | +2.18 pp |
Spider (out/full_spider/eval_heldout.json) |
1,350 | 43.63 % | 44.89 % | +1.26 pp |
The 10-database BIRD pilot (out/eval_heldout.json) reported a larger delta (+2.77 pp) at
smaller N; both are offline-proxy, not end-to-end EX numbers.
Contents
bird-train/<db_id>/ 52 dirs: fill_rules.py (when verified), meta.json,
verifier_report.json, attempts/attempt_N.py + report_N.json
spider-train/<db_id>/ 116 dirs, same shape
out/
sample.json 10-db pilot sample
sample_full_bird.json every eligible BIRD-train db + skip counts
sample_full_spider.json every eligible Spider-train db + skip counts
pairs.json, pairs_heldout.json pilot (10-db) pair sets
eval.json, eval_heldout.json pilot (10-db) two-arm reports
full_bird/{pairs,pairs_heldout,eval,eval_heldout}.json, run_all.txt
full_spider/{pairs,pairs_heldout,eval,eval_heldout}.json, run_all.txt
verification/ identity invariant, hit-rate, mechanism checks (t1-t6)
REPORT.md, experiments_subsection.tex
out/full_bird/ and out/full_spider/ are the multi-MB dumps the source git repo
deliberately does not commit (>1 MB rule in the package's own MANIFEST.md); this dataset
is their durable copy. Everything under bird-train/, spider-train/, out/sample.json,
out/pairs*.json, out/eval*.json and out/verification/ is also committed in the source
repo — mirrored here for a single place to pull the whole artefact set from.
Not included, by the source package's own contract (see its MANIFEST.md): generated
.sqlite instances, verifier trial databases, and anything under /dev/shm — all
transient and regenerable from fill_rules.py + the base databases.
Provenance / labels
- Label on every eval file:
offline-proxy— a proxy measurement (does the arm separate more near-miss pairs), not an end-to-end trained-model EX comparison. - No query in
fill_rules.pyconstruction is informed by RL rollouts; it operates on gold SQL and schema only. - Reproduce:
campaigns/chained_05b_tkde/valuefill_agents/scripts/run_all.sh(source repo); the second and later runs are free — every LLM call is cached bysha256(system + user + model + flags)underruns/<key>/.
Related CoRE-DB releases
thanhdath/CoRE-DB-BIRD-SPIDER— the frozen compact-database release this harness feeds.thanhdath/CoRE-DB-SQL-R1— the SQL-R1-matched CoRE-DB build.
- Downloads last month
- 12