# Fixed-prompt baseline inputs

Each `mini_dev_{sqlite,mysql,postgresql}.jsonl` contains 500 inputs aligned to
the current [`../data/`](../data/) files, question IDs 0–499. Send only the
row's `prompt` to a model. The separate `SQL` field is evaluation metadata,
not a model input. These files are prompts, not newly collected predictions
or benchmark scores.

The established fixed prompt wording and field order are unchanged. The
schema text for each database is reused byte-for-byte from the previous
published baseline and checked against all 33 published schema hashes.
Only the annotation fields and the question/evidence block have been
refreshed from the finalized Mini-Dev 1024 records. No old question or gold
SQL was used to fill the new annotations.

- [`prompt_template.txt`](prompt_template.txt): exact machine-rendered template.
- `schemas/`: the three pinned database-to-schema maps.
- [`validate_baseline_inference.py`](validate_baseline_inference.py): standalone
  checks for annotations, prompt rendering and pinned template/input/schema hashes.

Check offline from a downloaded repository snapshot, using Python
3.11 or newer. No hosted databases, model endpoints, keys or Dev release
checkout are required:

```bash
python baseline_inference/validate_baseline_inference.py
```

The validator needs only the files in this directory and the sibling `data/`
directory. The default Hugging Face configuration loads only
`data/`, not these expanded prompts. Join predictions by dialect and
`question_id`; use the original strict EX metric, not the cross-dialect
reference-comparison tolerance.