Datasets:
Reproducing SREB
Every number in this repository must be obtainable from data/ alone, without rerunning
any model. Rerunning a model is only needed to add a new model.
1. Check the dataset
python scripts/check_corpus.py --repo-root .
Verifies that every file listed in manifest.json matches its sha256, that every state
belongs to an existing family, that every refuted record carries refuting states and a
witness, and that every evidence file pins the current corpus and state-bank hashes. A
stale pin is a hard error: published evidence must never refer to inputs that are no
longer in the tree. CI runs exactly this command on every push.
2. Recompute the tables
python scripts/recompute_tables.py --repo-root .
Builds the summary tables (model × outcome, refutations per family) from data/evidence/.
The output must match docs/tables.md byte for byte; a mismatch means the published
tables are stale. CI reruns both the check and the regeneration and fails if the committed
file differs.
3. Cross-DBMS verification
docs/CROSS_DBMS.md records the PostgreSQL 16.15 re-run of the seed
families: the same corpus, states and saved answers, the corpus DDL executed unchanged,
1 008 candidate checks and zero status mismatches against the published MySQL evidence.
The raw check record is docs/cross_dbms_seed_check.json.
PostgreSQL is not yet first-class evidence in data/; the format stores one DBMS per
evidence file.
4. Model collection and evaluation
The scripts that collect model answers and execute candidates on a DBMS are not part of v0.2.0; the dataset publishes their outputs instead. Each record carries everything needed to audit it:
data/runs/<model>/<family>.json— model name, digest, quantization, engine version, temperature, seed (null when unset), context length, max tokens, the exact prompt and the raw response, plus the parsedcandidate_sql.data/evidence/<family>.json— DBMS and version, the corpus and state-bank pins, and per model: status, refuting states, the witness rows and the outcome of an independent replay in a fresh schema.data/states/states.json— the full rows of every allowed state, so a candidate can be re-executed against the same instances.
Re-running a model therefore requires reproducing the sampling protocol recorded in the run file and the state rows from the bank; the frozen hashes make a run comparable only if its inputs are identical.
What must never change silently
The protocol is pinned by hash. If the corpus, the state bank or the generator source changes, a new dataset version is required — the old runs stay valid for the old version and must not be mixed with new ones.
| Pinned | Where |
|---|---|
| dataset version | manifest.json, CITATION.cff, each file under data/ |
| corpus sha256 | manifest.json and protocol.corpus_sha256 of each evidence file |
| state bank sha256 | manifest.json and protocol.state_bank_sha256 of each evidence file |
| model digest | each file in data/runs/ |
| DBMS version | each file in data/evidence/ |
| sampling parameters | each file in data/runs/ |