thetmon/c1
Thetmon/c1 is a 4 billion parameter LoRA adapter fine-tuned from Qwen/Qwen3-4B-Instruct-2507 using QLoRA (4-bit, Unsloth). This adapter is specifically optimized to enhance structured output accuracy for formats like JSON, YAML, XML, TOML, and CSV. It improves the model's ability to generate precise and correctly formatted data structures by masking loss on intermediate reasoning. This model is ideal for applications requiring reliable structured data generation from a compact language model.
Loading preview...
Overview
The thetmon/c1 model is a LoRA adapter designed to enhance the structured output capabilities of the Qwen/Qwen3-4B-Instruct-2507 base model. It was fine-tuned using QLoRA (4-bit, Unsloth), making it an efficient addition for specific tasks. This repository provides only the adapter weights, requiring the base model to be loaded separately.
Key Capabilities
- Improved Structured Output: The primary objective of this adapter is to significantly boost the accuracy of structured data generation.
- Multi-format Support: Excels at producing outputs in various structured formats, including JSON, YAML, XML, TOML, and CSV.
- Efficient Fine-tuning: Utilizes QLoRA (4-bit) for efficient adaptation, making it practical for deployment.
- Focused Training: Loss is applied exclusively to the final assistant output, ensuring that the model's learning is concentrated on generating correct structured data, while intermediate Chain-of-Thought reasoning is masked.
Training Details
The adapter was trained on the u-10bei/structured_data_with_cot_dataset with a maximum sequence length of 2048 over 3 epochs. It used a learning rate of 2e-06 and LoRA parameters of r=64, alpha=128.
When to Use This Model
This adapter is particularly well-suited for applications where precise and reliable structured data generation is critical. If your use case involves extracting information into specific formats or generating configuration files, this model can provide a significant advantage by reducing errors in output structure.