docs/DATASET_CARD.md

Dataset card — SREB v0.3.0

Purpose

To make SQL rewrite quality measurable with executable evidence rather than declared success rates. A model that claims a rewrite preserves the result must survive the allowed states of the database; when it does not, the dataset records exactly where.

What the dataset contains

Item	Count
Query families	204 (68 library + 68 shop + 68 university)
Frozen states per family	28
Family–state pairs	5 712
Models with runs	6 (qwen2.5-coder:7b, qwen2.5:7b, qwen2.5:3b, phi3.5:3.8b, deepseek-coder-v2:latest, llama3.2:3b)
Run records (model × family)	1 224
Evidence records (model × family)	1 224
Executed checks (model × family × state)	34 272
Families with frozen states	204 (all)
Families with model runs	204 (all)

Coverage is complete for the corpus, the state banks and the model runs: every family has 28 frozen states, one answer per model and one evidence record per model. Six models are included. gemma4:26b is resource-blocked on the collection machine: the 25.2B Q4_K_M model is 17.3 GB, the machine has 15.7 GB of RAM and no GPU, and a single request did not complete within 18 minutes. A deterministic rules baseline is not published because the rule module in the source application is an analyser, not a rewrite generator; it produces findings and suggestions rather than executable candidate SQL.

The primary execution engine is MySQL 8.4.11. The evaluation mechanism was additionally cross-checked on PostgreSQL 16.15 for the 12 seed families (1 008 candidate checks, no status mismatches, no DDL changes); see CROSS_DBMS.md. PostgreSQL evidence is not yet first-class in the dataset format; the university schema has not been ported to PostgreSQL yet.

Structure

corpus/families.json, corpus/shop.json and corpus/university.json

One file per schema. Each declares the schema DDL, the constraints it enforces, and the query families with a query-shape profile and a designed breaking_profile. profile is derived from the reference SQL only, never from model answers; breaking_profile is the state profile the family is designed to be challenged by (measured refutations live in evidence/).

{
  "schema": "library",
  "ddl": ["CREATE TABLE lib_readers (...)", "..."],
  "constraints": ["lib_readers.id PRIMARY KEY", "lib_loans.reader_id REFERENCES lib_readers(id)"],
  "families": [
    {"id": "loan_count", "schema": "library", "reference_sql": "SELECT ...",
     "profile": "outer_join_aggregate", "breaking_profile": "empty",
     "intent": "Число выдач, включая ноль", "ordered": true}
  ]
}

shop.json has the same shape for the shop schema (categories, products, customers, orders, order_items, payments, reviews) with PK, FK, NOT NULL and CHECK constraints on non-negative amounts and quantities and on enumerated columns.

states/states.json

Two frozen banks, one per schema, each of 28 states: 8 designed profiles (fixed, empty, no_rows, multiplicity, zero_value, nullable, boundary, ties) and 20 pseudo-random states with fixed seeds 0…19. A bank is applied to every family of its schema and expanded one record per (family, state) pair, so each record is self-contained:

{"family": "loan_count", "profile": "multiplicity_0", "profile_family": "multiplicity",
 "seed": 0, "rows": {"lib_readers": [[...]], "lib_books": [[...]], "lib_loans": [[...]]}}

profile is the stable state id used by evidence (fixed_0, empty_0, no_loans_0, multiplicity_0, zero_fee_0, nullable_0, boundary_0, ties_0, random_uniform_0 … random_uniform_19); profile_family groups the ids (fixed, empty, no_loans, multiplicity, zero_fee, nullable, boundary, ties, random_uniform). Random states carry the generator seed.

The library bank is the frozen seed bank: its domain-named no_loans_0 and zero_fee_0 implement the generic no_rows and zero_value profiles. The shop bank was generated deterministically (random.Random(1000 + seed)), and every state was loaded into MySQL in tables with real constraints — foreign keys and CHECK constraints enabled — before any query was executed on it. A state that violates the schema is not part of the dataset.

runs//.json

One file per model and family. The model tag is in model; the directory name is a filesystem-safe slug (a colon is not portable in a path).

{"family": "loan_count", "model": "qwen2.5-coder:7b", "model_digest": "dae161e2…",
 "quantization": "Q4_K_M", "parameter_size": "7.6B",
 "protocol": {"engine": "ollama", "engine_version": "0.40.4", "temperature": 0,
              "seed": null, "num_ctx": 4096, "num_predict": 1200},
 "prompt": "...", "raw_response": {...}, "response_text": "...",
 "candidate_sql": "...", "markdown_removed": false, "provenance": {...}}

seed: null means the run protocol did not specify a seed; the engine default was used. It is recorded as null rather than invented. raw_response is the engine response verbatim, including token-level fields.

evidence/.json

One file per family with one evidence record per model. Outcomes:

Status	Meaning
refuted	at least one state breaks the rewrite; refuting_states and witness are present
preserved	the candidate differs from the reference SQL and no state broke it (not a proof of equivalence)
no_rewrite	the model returned the reference query unchanged
unsupported	the candidate could not be parsed or used unsupported constructs
error	the reference/candidate could not be executed or no response was stored

{"family": "union_bag", "reference_sql": "...",
 "protocol": {"dbms": "mysql", "dbms_version": "8.4.11", "corpus_sha256": "...",
              "state_bank_sha256": "..."},
 "evidence": [
   {"model": "qwen2.5:7b", "model_digest": "...", "status": "refuted",
    "candidate_sql": "...", "refuting_states": ["multiplicity_0", "nullable_0", "ties_0"],
    "witness": {"state": "multiplicity_0", "state_profile": "multiplicity",
                "reference_rows": [[1], [1], [1], [1]],
                "candidate_rows": [[1], [1], [1]],
                "interface_equal": true, "independent_replay": "counterexample"},
    "checks": 28}
 ]}

Rows are exactly what the DBMS returned (a list of rows, a row is a list of column values). independent_replay records that the refuting pair was executed again in a fresh private schema.

The file protocol pins its inputs: protocol.corpus_sha256 is the sha256 of the corpus file of the evidence file's schema (data/corpus/families.json for the library schema, data/corpus/shop.json for the shop schema), and protocol.state_bank_sha256 is the sha256 of data/states/states.json. check_corpus.py recomputes both and rejects a file whose pin points to inputs that are no longer in the tree.

manifest.json

SHA-256 of every file under data/, plus the recomputable counts, the model digests and the provenance of the upstream runs. A file that is not listed is a validation error, so no data can escape the frozen hashes.

manifest.counts is the machine-readable statement of coverage; check_corpus.py recomputes every value from data/. manifest.staging records the current stage of the dataset — for v0.3.0 it is university-runs: the corpus has 204 families on three schemas, the frozen state banks cover all 204 and model runs and evidence also cover all 204; the PostgreSQL cross-check covers the 12 seed families (see CROSS_DBMS.md). The same note is repeated in the manifest and the dataset card so that coverage claims stay machine-checkable.

Integrity rules

1. Every file under data/ must be listed in manifest.json with its sha256; a mismatch means the dataset was modified after freezing.
2. manifest.counts must be recomputable from data/ alone.
3. status: refuted requires at least one state in refuting_states and a non-empty witness whose state is one of them; preserved/no_rewrite must not carry states.
4. status: preserved means no refuting state was found — it does not prove equivalence.
5. Models are described by digest, not by name alone. Two runs of "the same" tag can differ.
6. Every evaluated family must have a run record for every model directory and an evidence file. A run record may only exist for a family that has frozen states.
7. Every evidence file must pin the current corpus and state-bank hashes in its protocol; check_corpus.py recomputes them, so CI fails when evidence and inputs drift apart.

Known limitations

• Synthetic schemas: coverage of real-world query shapes is partial.
• A finite state bank is not a proof; preserved is evidence of absence, not absence of bugs.
• Model behaviour changes with engine version; the engine version is part of the protocol.
• One response per (model, family), temperature 0, no explicit seed: the raw text is stored as produced but a rerun is not guaranteed to be byte-identical.
• Candidate rewrites were executed on MySQL only; the PostgreSQL cross-check covers the 12 seed families and is not first-class evidence. The university schema has not been ported to PostgreSQL yet.
• The state bank is shared by all families of a schema: it stresses the schema, not every individual query shape. Coverage of edge cases that are specific to one family is a design target, not a guarantee.
• breaking_profile is a design intent, not a measurement.
• The model set is six models; the originally planned gemma4:26b did not fit the collection hardware (17.3 GB model on 15.7 GB RAM, no GPU) and is excluded. Engine versions span Ollama 0.40.4–0.40.6 because the series were collected on different days; the version is recorded in every run and evidence protocol.
• No deterministic rules baseline: the rule module of the source application is an analyser, not a rewrite generator.
• Service-level failures (empty or failed responses) are recorded as error and are not retried, because the protocol allows one answer per family: 33 of 1 224 records, mostly qwen2.5:3b (16) and llama3.2:3b (10).
• 23 answers are unparsable or invalid SQL and are stored as unsupported (15 phi3.5:3.8b, 6 qwen2.5:3b, 1 qwen2.5-coder:7b, 1 llama3.2:3b); execution failures caused by the candidate SQL itself (missing derived-table aliases, ONLY_FULL_GROUP_BY violations, unknown columns) are error records, not silently dropped.