exp-0216-005-db-balanced-qwen2.5-7b

Fine-tuned from Yano/exp-0212-001-alfworld-qwen2.5-7b (001 ALFWorld SFT model) using QLoRA (4-bit, Unsloth).

Purpose

DB Bench training with balanced data (v1-v4 mixed, INSERT/UPDATE downsampled). Addresses 004's SELECT degradation (76.5% -> 41.0%) caused by INSERT/UPDATE data imbalance.

Data Strategy

• Combined all DB Bench v1-v4 (3,060 total)
• Downsampled INSERT to 200, UPDATE to 150
• Kept all SELECT/query-type samples
• Final dataset: ~1470 samples (INSERT+UPDATE ~24.5%)

Training Configuration

• Base model: Yano/exp-0212-001-alfworld-qwen2.5-7b
• Method: QLoRA (4-bit), merged to 16-bit
• Max sequence length: 2048
• Epochs: 3
• Learning rate: 2e-05
• LoRA: r=64, alpha=128
• Batch size: 2 (grad accum 16)
• Warmup ratio: 0.1
• Collator: AllAssistantTurnsCollator (all turns supervised)