Model Overview

This model, kiratan/qwen3-4b-structeval-lora-57-merged, is a specialized version of the Qwen3-4B-Instruct-2507 base model, fine-tuned to significantly improve its performance on structured output tasks. It was developed by kiratan using a multi-step QLoRA training process, with the LoRA weights merged directly into the base model for ease of deployment without requiring PEFT.

Key Capabilities

• Optimized Structured Output: Specifically trained to generate accurate JSON, YAML, XML, TOML, and CSV formats.
• Targeted Loss Application: During training, loss was applied only to the final assistant output, masking intermediate reasoning (Chain-of-Thought) to focus on direct structured generation.
• Robust Training Regimen: Underwent a multi-step training process using various daichira/structured-mix-sft datasets, including a "hard" dataset for increased robustness.
• Specific Format Upsampling: Certain difficult conversions, such as csv → json and xml → yaml, were upsampled during training to enhance proficiency.

Good For

• Applications requiring reliable and accurate generation of structured data (e.g., configuration files, API responses).
• Tasks involving conversion between different structured data formats.
• Developers looking for a pre-merged model that can be loaded directly for structured output without additional PEFT setup.