qwen3-4b-h100-mix-4k

This LoRA adapter is optimized for structured output (JSON/XML/YAML/CSV) using Qwen3-4B-Instruct-2507 and Unsloth. Trained on H100 with merged datasets (5k-mix + hard-4k) and a context length of 4096.

Training Configuration

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Max sequence length: 4096
• Epochs: 1
• Learning rate: 5e-5
• Batch size (effective): 64 (BS=16, GradAccum=4)