ujwal00/qwen3-1.7b-json
The ujwal00/qwen3-1.7b-json model is a 2 billion parameter Qwen3-1.7B fine-tune, specifically optimized for native JSON-Schema compliant output. It significantly improves performance on complex, nested JSON schemas without requiring grammar-constrained decoding. This model excels at generating structured JSON data that strictly adheres to a given JSON Schema, making it ideal for applications requiring reliable and validated JSON outputs.
Loading preview...
Qwen3-1.7B-JSON: Native JSON-Schema Compliance
This model is a fine-tuned version of Qwen/Qwen3-1.7B (2 billion parameters, 32768 tokens context length) designed to generate JSON output that strictly conforms to a provided JSON Schema. Its key differentiator is its ability to achieve high schema compliance natively, without relying on external grammar-constrained decoding engines, particularly for complex and nested schemas.
Key Capabilities & Performance
- Enhanced JSON-Schema Compliance: Achieves significant gains in native schema compliance, especially on challenging nested (
$ref,$defs) schemas, improving from 58% to 72–74% on held-out data. - Robust on Complex Schemas: Demonstrates substantial improvements across various datasets from JSONSchemaBench, with gains of +13% to +40% on 'medium' to 'hard' categories, including Github_hard and Snowplow.
- Low Parse-Fail Rate: Maintains a low parse-fail rate of 2.5–4%, indicating reliable JSON generation.
- QLoRA Fine-tuning: Trained using QLoRA SFT on approximately 4k schema-instance pairs, ensuring robust performance on validated targets.
Ideal Use Cases
- Structured Data Generation: Perfect for applications requiring precise, validated JSON outputs from a language model.
- Function Calling & API Interaction: Can be used to generate API requests or responses that adhere to predefined JSON schemas.
- Small Model Deployment: Lifts native compliance for small models in high-volume, structured tasks, making them more usable without additional decoding engines.