README.md

metadata

license: cc-by-sa-4.0
pretty_name: Tiny Aya SQL agent - BIRD Mini-Dev evaluation trajectories
language:
  - en
tags:
  - text-to-sql
  - bird
  - evaluation
  - trajectories
  - agent
configs:
  - config_name: bird_scores
    default: true
    data_files:
      - split: base
        path: runs/final-bird-sql-20261109-02/bird-base/metrics.jsonl
      - split: a1500
        path: runs/final-bird-sql-20261109-02/bird-a1500/metrics.jsonl
      - split: b
        path: runs/final-b-bird-sql-20261110-01/bird-b/metrics.jsonl

Tiny Aya SQL agent: BIRD Mini-Dev evaluation records

Per-question scores and full agent trajectories for the untuned Tiny Aya Global, A1500 and B on 493 BIRD Mini-Dev questions (SQLite).

• Use: BIRD was used for the final evaluation only. It was never used for training, tuning or choosing a model.

• Licence: this dataset is kept separate because BIRD's licence is share-alike.

• Models: B and A1500

• Main data and logs: satish9177/tiny-aya-sql-agent-data

• Code: github.com/satish9177/aya-db-agent

• Write-up: Teaching Tiny Aya to query databases (HF Article)

Scores (strict; 493 of 500 questions)

Model	Strict	Set
Tiny Aya Global, untuned	0 (0.0%)	-
A1500	8 (1.6%)	8
B (submitted)	7 (1.4%)	8

• Exclusions: seven questions were excluded before any model use, because their reference query exceeded the execution cap. They are listed in eval-inputs/data/bird/final-v1/exclusions.jsonl.
• Main failure: the models mostly run out of their eight tool steps on BIRD's large schemas.

Contents

Path	Contents
eval-inputs/data/bird/final-v1/tasks.jsonl	the 493 questions in our tool format, with BIRD's evidence notes where present, reference SQL and recomputed gold rows
runs/final-bird-sql-20261109-02/	Base and A1500: scores, trajectories, submissions, AuditKIT verdicts, job logs
runs/final-b-bird-sql-20261110-01/	B: the same
MANIFEST.json	SHA256 and size of every file

Databases are not included. Get them from the official source (see below), then place them where the task files point.

About these records

• Edits for publication: personal details were removed after the runs.
  • hf-user stands for the HF account a job ran on.
  • Job links are reduced to job IDs.
  • Hashes inside the records refer to the files as they were when each job ran.
• Sample rows: they are shown when the agent describes a table, and they come from BIRD's public databases unchanged. One of them is an email address in a public user profile.
• Difficulty labels: the labels come from the source file as-is. That file labels 460 of 500 questions "challenging", unlike BIRD's published split, so we report BIRD only overall.

Licence and credits

• Licence: CC BY-SA 4.0, as these records are derived from BIRD Mini-Dev.
• Model outputs: the records also contain outputs of Tiny Aya Global (Cohere Labs, CC-BY-NC-4.0).
• Source: birdsql/bird_mini_dev, revision 5b256b5a9c17478f06ed6912612e9a14d17065c8. The databases come from the card-linked official SQLite images.

Please cite BIRD:

@inproceedings{li2023bird,
  title     = {Can LLM Already Serve as A Database Interface? A BIg Bench for Large-Scale Database Grounded Text-to-SQLs},
  author    = {Li, Jinyang and Hui, Binyuan and Qu, Ge and others},
  booktitle = {Advances in Neural Information Processing Systems 36},
  year      = {2023}
}