Qwen2.5-Coder-7B Text-to-SQL — filtered-train variant (QLoRA)
Same recipe as junmingg/qwen2.5-coder-7b-text2sql,
but the 80 training rows whose gold SQL is unparseable (0.32% — WikiSQL label noise) were dropped before
fine-tuning (24,920 train rows). This is the label-noise ablation described in the main model card.
Results (held-out 500 examples — identical references)
| Model | Exact match | Semantic equiv. | SQL validity |
|---|---|---|---|
| Base (zero-shot) | 3.8% | 67.0% | 100.0% |
| Main model (full 25k train) | 78.8% | 86.2% | 99.2% |
| This model (filtered 24.9k train) | 78.6% | 85.4% | 99.8% |
On the valid-reference subset (497 rows) this model reaches 100% validity. Cleaning the noisy labels moved validity, not accuracy — exact-match and semantic-equivalence are unchanged within run-to-run noise.
Use the main model unless you specifically want this ablation. Usage, training details, evaluation methodology, and limitations are identical to the main model card. Code + harness: https://github.com/junmingg/Unsloth-Qwen2.5-Coder-7b-Text-to-SQL-SFT
License & attribution
Apache-2.0 (base model). Training data: b-mc2/sql-create-context (CC-BY-4.0).