kiratan/qwen3-4b-structeval-lora-57-merged
kiratan/qwen3-4b-structeval-lora-57-merged is a 4 billion parameter language model based on Qwen3-4B-Instruct-2507, fine-tuned specifically for enhanced structured output generation. This model excels at producing accurate JSON, YAML, XML, TOML, and CSV formats. It was trained using QLoRA and merged into the base weights, making it ready for direct use in applications requiring reliable structured data extraction or generation.
Loading preview...
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-sftdatasets, including a "hard" dataset for increased robustness. - Specific Format Upsampling: Certain difficult conversions, such as
csv → jsonandxml → 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.