Create README.md

README.md

@@ -0,0 +1,88 @@
+---
+license: apache-2.0
+task_categories:
+  - text-generation
+language:
+  - pt
+tags:
+  - text-to-sql
+  - portuguese
+  - benchmark
+  - sql
+  - postgis
+  - geospatial
+  - environmental-registry
+size_categories:
+  - n<1K
+---
+
+# Environmental Registry Test Set
+
+This dataset is the anonymized primary benchmark used for evaluating agentic Portuguese Text-to-SQL over a real PostgreSQL/PostGIS environmental-registry database. The underlying production database is not released, but the benchmark metadata and gold labels are provided for transparency and comparison.
+
+Code and reproducibility repository:
+
+https://github.com/Boakpe/distilled-slms-for-text-to-sql-pt-br
+
+Related collection:
+
+https://huggingface.co/collections/Boakpe/distilled-slms-for-text-to-sql-pt-br
+
+## Dataset Summary
+
+- Rows: 180
+- Split: `test`
+- SQL questions: 90
+- Clarification questions: 45
+- Unanswerable questions: 45
+- SQL difficulty split: 20 easy, 30 medium, 30 hard, 10 expert
+- Language: Brazilian Portuguese
+- Database type: PostgreSQL/PostGIS environmental-registry database
+
+The primary schema contains 19 tables and covers rural properties, municipalities, Indigenous Lands, conservation units, environmental embargoes, settlements, quilombola territories, federal public lands, deforestation events, and spatial-overlap facts.
+
+## Columns
+
+- `id`: benchmark identifier.
+- `pergunta`: natural-language question in Portuguese.
+- `sql`: gold SQL for SQL questions, or terminal label such as `clarification` / `unanswerable` for non-SQL questions.
+- `dificuldade`: SQL difficulty label.
+- `categorias`: topic labels.
+- `order_matters`: whether row order is semantically required.
+- `num_column_matters`: whether exact column count is required.
+- `num_row_matters`: whether exact row count is required.
+
+## Evaluation Design
+
+The benchmark combines:
+
+- SQL generation.
+- Ambiguity detection through clarification questions.
+- Unanswerability detection.
+- PostGIS/geospatial reasoning.
+- Temporal filters, joins, aggregations, and domain-specific value grounding.
+
+SQL evaluation uses execution comparison rather than exact SQL-string match. The reported primary metric is relaxed execution accuracy, with strict execution accuracy also reported as a diagnostic.
+
+## Reported Results
+
+Primary environmental-registry benchmark, Pass@1:
+
+| Model | Overall | Strict SQL | Relaxed SQL | Non-SQL | Clarification | Unanswerable |
+|---|---:|---:|---:|---:|---:|---:|
+| DeepSeek V4 Pro | 86.7 | 44.4 | 83.3 | 90.0 | 93.3 | 86.7 |
+| GLM 5.1 | 84.4 | 44.4 | 80.0 | 88.9 | 86.7 | 91.1 |
+| Qwen3-4B-Thinking FT | 78.9 | 34.4 | 70.0 | 87.8 | 86.7 | 88.9 |
+| Qwen3.5-27B-Q3_K_M teacher | 75.0 | 40.0 | 70.0 | 80.0 | 75.6 | 84.4 |
+| Qwen3-4B-Thinking-2507 base | 56.1 | 28.9 | 36.7 | 75.6 | 71.1 | 80.0 |
+
+Fine-tuned model Pass@K:
+
+| Setting | Overall | Relaxed SQL | Non-SQL |
+|---|---:|---:|---:|
+| Pass@1 | 78.9 | 70.0 | 87.8 |
+| Pass@5 | 91.7 | 87.8 | 95.6 |
+
+## License
+
+Apache 2.0.