SchemaSage-SQL Qwen3 4B Longer QLoRA Adapter
Experimental 200-step QLoRA adapter for SchemaSage-SQL. This is an evaluation-stage artifact, not a production release.
Training
- Base model:
Qwen/Qwen3-4B-Instruct-2507 - Train rows: 2,048
- Trainer eval rows: 128
- Optimizer steps: 200
- Hardware: Hugging Face Jobs
a10g-large - Job ID:
6a0ccb203aba298b21d143ad
Held-Out Evaluation
64 held-out examples from gretelai/synthetic_text_to_sql test.
| Metric | Value |
|---|---|
| Exact match | 0.2812 |
| Normalized exact match | 0.2969 |
| SQL parse validity | 1.0000 |
| Schema adherence rate | 0.9219 |
| Hallucinated table rate | 0.0312 |
| Hallucinated column rate | 0.0781 |
| Unsafe query rate | 0.0938 |
| Execution accuracy | 0.8750 |
| Execution comparable examples | 40 |
| Mean generated SQL length | 13.34 |
| Mean latency seconds | 5.99 |
Evaluation Summary
The longer run improves exact match and execution accuracy over the 10-step smoke adapter, but it is not release-ready because unsafe query rate regressed. The training data includes destructive SQL targets, and the model learned to imitate them. Next stage should clean or transform unsafe examples into refusals before additional training.
See reports/longer_adapter_eval_analysis.md for detailed findings.
- Downloads last month
- 4
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for rishhh/schemasage-sql-qwen3-4b-longer-200
Base model
Qwen/Qwen3-4B-Instruct-2507