---
license: cc-by-4.0
language:
- en
tags:
- text-to-sql
- schema-linking
pretty_name: GRAST-SL evaluation sets
configs:
- config_name: spider_dev
  data_files: evaluation_set/spider_dev.csv
- config_name: bird_dev
  data_files: evaluation_set/bird_dev.csv
- config_name: spider2_lite_256
  data_files: evaluation_set/spider2_lite_256.csv
- config_name: spider2_snow_256
  data_files: evaluation_set/spider2_snow_256.csv
- config_name: spider_dev_case_by_case
  data_files: comparison_with_prior_golds/spider_dev_case_by_case.parquet
- config_name: bird_dev_case_by_case
  data_files: comparison_with_prior_golds/bird_dev_case_by_case.parquet
- config_name: spider_dev_label_errors
  data_files: comparison_with_prior_golds/spider_dev_label_errors.parquet
- config_name: bird_dev_label_errors
  data_files: comparison_with_prior_golds/bird_dev_label_errors.parquet
- config_name: summary_statistics
  data_files: comparison_with_prior_golds/summary_statistics.parquet
---

# GRAST-SL evaluation sets

Schema-linking evaluation sets of the paper *"Scaling Text2SQL via LLM-efficient Schema Filtering
with Functional Dependency Graph Rerankers"*.

## Structure

```
evaluation_set/
  spider_dev.csv         1,034 questions - question, gold SQL, gold columns
  bird_dev.csv           1,534 questions
  spider2_lite_256.csv     256 instances (Spider 2.0-Lite)
  spider2_snow_256.csv     256 instances (Spider 2.0-Snow)
  check_gold.py          gold-error checker: every gold set non-empty and
                         every gold column present in the database schema
                         (all four sets pass with 0 errors)
schemas/                 full schema list per database (db_id -> columns);
                         join on each row's db_id to get the schema of that
                         sample; what check_gold.py validates against
comparison_with_prior_golds/
  {spider,bird}_dev_case_by_case.parquet   per-question comparison with the prior
                                           released gold labels (lexical matching);
                                           error columns first
  {spider,bird}_dev_label_errors.parquet   one row per wrong gold column + reason
  summary_statistics.parquet               aggregate numbers
```


## How the gold was built

One pipeline for all three benchmarks:

- **Spider / BIRD**: released gold SQL + schema → **o4-mini** extracts used columns →
  post-processing: gold ⊆ schema (0 invalid) → no empty gold set → parser cross-check →
  PK convention for star-only tables → per-instance audit.
- **Spider 2.0 (Lite / Snow)**: shard-table grouping (`events_20201101`, … → `events_*`) →
  same steps as above.

`check_gold.py` re-verifies gold ⊆ schema and non-emptiness from the shipped files. Prior
released Spider/BIRD golds used lexical name matching (a name is credited to every table
containing it) — the case-by-case CSVs list each resulting error.

## Citation

```bibtex
@article{hoang2025scaling,
  title={Scaling Text2SQL via LLM-efficient Schema Filtering with Functional Dependency Graph Rerankers},
  author={Hoang, Thanh Dat and Nguyen, Thanh Tam and Huynh, Thanh Trung and Yin, Hongzhi and Nguyen, Quoc Viet Hung},
  journal={arXiv preprint arXiv:2512.16083},
  year={2025}
}
```

Code: https://github.com/thanhdath/grast-sql