daichira/qwen3-4b-structeval-V-1228-lora

qwen3-4b-structeval-lora

This is a LoRA adapter finetuned on Qwen3-4B-Instruct-2507 for structured code generation tasks.

Model Description

• Base Model: unsloth/Qwen3-4B-Instruct-2507
• Finetuning Technique: QLoRA (via Unsloth)
• Objective: Generating pure structured code (HTML, SVG, Angular, etc.) without explanations or markdown fences.

Training Details

• Dataset: Stratified SFT dataset (1,500 samples)
• Task Types: HTML, Angular, React, SVG
• Training Config:
  • Learning Rate: 2e-4
  • Batch Size: 16 (Global)
  • Epochs: 1
  • Optimizer: Paged AdamW (32-bit)

Intended Use

This model is specialized for generating structured code. It is designed to output ONLY the code.

License

This model follows the license of the base model (Qwen License).