kikansha-Tomasu/Qwen3-4B-Instruct-2507-sft1
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:Feb 6, 2026License:apache-2.0Architecture:Transformer Open Weights Cold
The kikansha-Tomasu/Qwen3-4B-Instruct-2507-sft1 is a 4 billion parameter instruction-tuned language model, fine-tuned from Qwen/Qwen3-4B-Instruct-2507 using QLoRA. This model is specifically optimized for generating accurate structured outputs such as JSON, YAML, XML, TOML, and CSV. It features a 32768 token context length and is designed for direct use with merged full model weights.
Loading preview...
Overview
This model, kikansha-Tomasu/Qwen3-4B-Instruct-2507-sft1, is a 4 billion parameter instruction-tuned language model. It is a merged model, fine-tuned from the Qwen/Qwen3-4B-Instruct-2507 base using QLoRA (4-bit, Unsloth). The full model weights are included, allowing for direct use without needing to load the base model separately.
Key Capabilities
- Structured Output Generation: The primary objective of this fine-tuning was to enhance the model's accuracy in generating structured data formats.
- Supported Formats: Excels at producing outputs in JSON, YAML, XML, TOML, and CSV.
- Efficient Training: Utilized QLoRA (4-bit) for efficient fine-tuning, with a focus on applying loss only to the final assistant output while masking intermediate reasoning (Chain-of-Thought).
Training Details
- Base Model: Qwen/Qwen3-4B-Instruct-2507
- Method: QLoRA (4-bit)
- Max Sequence Length: 512
- Epochs: 1
- Learning Rate: 1e-06
- LoRA Configuration: r=64, alpha=128
- Training Data:
u-10bei/structured_data_with_cot_dataset_512_v2(MIT License)
Good For
- Applications requiring precise and reliable generation of structured data.
- Tasks where output format consistency (e.g., API responses, configuration files) is critical.