Nicholas55555/qwen2.5-1.5b-weather-intent-merged
Nicholas55555/qwen2.5-1.5b-weather-intent-merged is a 1.5 billion parameter Qwen2.5-Instruct model, fine-tuned and merged with a LoRA adapter for weather intent parsing. This model is specifically designed to convert natural language weather questions into structured JSON intent objects. It offers a 32768 token context length and is optimized for direct serving or quantization to GGUF formats. The model excels at accurately extracting and structuring weather-related information from user queries.
Loading preview...
Overview
Nicholas55555/qwen2.5-1.5b-weather-intent-merged is a specialized 1.5 billion parameter model built upon the Qwen2.5-1.5B-Instruct architecture. It integrates a LoRA adapter, specifically trained for weather intent parsing, directly into the base model. This merged configuration allows for efficient deployment, including quantization to GGUF formats for use with tools like llama.cpp or Ollama.
Key Capabilities
- Structured Intent Parsing: Converts natural language weather questions into a compact, structured JSON intent.
- High Accuracy: Achieves high performance on intent extraction, with a 98.6% exact match and 99.7% field accuracy on held-out evaluation data, significantly outperforming the base model's few-shot performance in a zero-shot setting.
- Efficient Deployment: Provided as a merged fp16 model, suitable for direct serving or further quantization.
Use Cases
This model is ideal for applications requiring precise extraction of weather-related information from user input, such as:
- Voice Assistants: Powering weather query capabilities in smart speakers or mobile assistants.
- Chatbots: Enabling chatbots to understand and respond to complex weather questions.
- Data Processing: Structuring unstructured weather queries for downstream analysis or API calls.
Performance Highlights
| metric | base | finetuned |
|---|---|---|
| valid JSON | 100.0% | 100.0% |
| exact match | 64.5% | 98.6% |
| field accuracy | 90.7% | 99.7% |
| slot F1 | 0.894 | 0.996 |
These results demonstrate a substantial improvement in intent parsing accuracy compared to the base model, particularly in exact match and field accuracy, achieved with zero-shot inference after fine-tuning.