dataslab/DLM-NL2JSON-4B

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Mar 19, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

DLM-NL2JSON-4B by Data Science Lab., Ltd. is a 4-billion parameter LoRA-merged Qwen3-4B model specifically fine-tuned for structured JSON extraction from Korean natural language queries. This domain-specific language model (DLM) significantly outperforms larger general-purpose models like GPT-4o and Qwen3.5-35B on its specialized task, achieving 94.4% accuracy. It is designed for converting Korean public/economic data queries into predefined JSON schemas, particularly for the Busan Metropolitan City public data analytics service.

Loading preview...

DLM-NL2JSON-4B: Specialized Korean NL-to-JSON Extraction

DLM-NL2JSON-4B is a 4-billion parameter model developed by Data Science Lab., Ltd., built upon a LoRA-merged Qwen3-4B base. It is a domain-specific language model (DLM) engineered for a singular purpose: converting Korean natural language queries into structured JSON outputs, specifically for the Busan Metropolitan City public data analytics service.

Key Capabilities & Performance

  • Exceptional Accuracy: Achieves 94.4% accuracy (96.8% adjusted) on structured JSON extraction from Korean queries, significantly outperforming GPT-4o (+14%p) and Qwen3.5-35B (+22%p) on its specific task.
  • Task-Specific Excellence: Dominates 8 out of 10 evaluated task categories, showing substantial gains (e.g., +43%p vs GPT-4o on ALP tasks).
  • Efficient: Despite its small size (4B parameters), it demonstrates that highly specialized fine-tuning can yield superior results compared to much larger, general-purpose models for constrained tasks.
  • Korean Language Support: Trained exclusively on Korean data and prompts for public/economic data analysis.

Important Considerations

  • Service-Specific: This model is not a general-purpose NL-to-JSON converter. It is strictly designed for a fixed set of predefined schemas and requires specific system prompts and special tokens (<TASK_CSM>, etc.) for correct operation.
  • Intended Use: Ideal for applications requiring precise, structured data extraction from Korean natural language within a predefined domain, such as public data analytics dashboards.
  • Limitations: Weaker performance on complex CPI category (86.3%) and not suitable for arbitrary JSON schemas or general chatbot interactions.

Finetunes
1 models