Overview

Yano/exp-0216-005-db-balanced-qwen2.5-7b is a 7.6 billion parameter model fine-tuned from Yano/exp-0212-001-alfworld-qwen2.5-7b using QLoRA (4-bit, Unsloth). Its primary purpose is to improve performance on database benchmark tasks by addressing data imbalance, specifically the degradation in SELECT query performance observed in previous iterations due to an overrepresentation of INSERT/UPDATE data.

Key Capabilities

• Optimized for Database Operations: Specifically trained to handle various database interactions, including SELECT, INSERT, and UPDATE queries.
• Balanced Data Handling: Utilizes a unique data strategy that combines DB Bench v1-v4 data while downsampling INSERT and UPDATE samples to ensure a more balanced representation, preventing performance degradation on query-type tasks.
• QLoRA Fine-tuning: Leverages QLoRA (4-bit) for efficient fine-tuning, with the final model merged to 16-bit.

Training Details

The model was trained for 3 epochs with a learning rate of 2e-05, using a maximum sequence length of 2048. The LoRA configuration included r=64 and alpha=128, with all assistant turns supervised during training.