Fine-tuned from Yano/exp-0216-005-db-balanced-qwen2.5-7b (005 DB Bench balanced model) using QLoRA (4-bit, Unsloth).
Purpose
INSERT recovery training on top of 005.
005 achieved SELECT 49.2%, INSERT 31.9%, UPDATE 80.0% (overall_cat 53.7%).
INSERT is the biggest bottleneck — this experiment increases INSERT data to 300 samples
with a gentle learning rate (1e-5) and 1 epoch to recover INSERT without breaking other metrics.
Data Strategy
Combined all DB Bench v1-v4 (3,060 total)
Downsampled INSERT to 300 (005: 200 → 028: 300), UPDATE to 150
Kept all SELECT/query-type samples
Final dataset: ~1570 samples
Training Configuration
Base model: Yano/exp-0216-005-db-balanced-qwen2.5-7b (005, DB ability already acquired)
