---
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags: [text-to-sql, grpo, core-db]
---
# FINER-SQL-3B-v2

FINER-SQL 3B SQL-Writer continually trained with GRPO (format + execution reward) on **CoRE-DB**, the compact
counterexample-enriched database environment of *Small Language Models for Text-to-SQL against Spurious Correctness*.
Checkpoint step 1000 (dev-selected). Code: [github.com/thanhdath/core-db](https://github.com/thanhdath/core-db).

| BIRD dev EX | Spider dev EX | Spider dev TS | Spider test EX |
|---:|---:|---:|---:|
| 68.84 (1056/1534) | 86.75 (897/1034) | 81.53 (843/1034) | 87.66 (1882/2147) |

k = 30 samples, T = 1.0, seed 42, at most 2048 new tokens, vote-after-verify selection, official evaluators (BIRD:
evaluation_bird_ex.py with SQLite 3.45.3 and `--meta_time_out 3600`; Spider: taoyds/test-suite-sql-eval @e97acc54,
default flags; TS on the distilled Spider dev test suites). Prompts: `thanhdath/core-db-evaluation-prompts`
(`bird_dev`, `spider_dev`; Spider test: `spider_test_full_schema`).

Provenance: these weights are byte-identical to `thanhdath/3b-core-noam-ict` @ `0c74daf264c89d9e4bdf1378bcf065cb76b1ef0f`, sub-folder `checkpoints/step01000/checkpoint-1000` (training step 1000).
The saved predictions and the re-scoring script are in `scripts/rq2/` of the code repository.