chyungwon/police-report-analysis-model
The chyungwon/police-report-analysis-model is a 1 billion parameter language model developed by chyungwon, fine-tuned from Google's Gemma-3-1B-it. It specializes in analyzing Korean crime incident reports, performing logical reconstruction of events (Context Generation) and accurate classification of incident types (Kind Classification). This model is optimized for specialized forensic text analysis tasks, leveraging QLoRA for efficient training and deployment.
Loading preview...
Overview
This model, developed by chyungwon, is a specialized 1 billion parameter language model built upon Google's Gemma-3-1B-it. It is specifically fine-tuned for the analysis of Korean crime incident reports, offering unique capabilities for forensic text processing.
Key Capabilities
- Context Generation: Analyzes incident reports to logically reconstruct the situation and factual relationships of a crime.
- Kind Classification: Accurately classifies the type of incident based on the reconstructed context.
Training Details
The model was fine-tuned using the QLoRA (Quantized Low-Rank Adaptation) technique, ensuring high memory efficiency and fast training. Key training parameters include:
- Quantization: 4-bit NF4 quantization (BitsAndBytes).
- PEFT: LoRA applied with Rank r=16 and alpha=32.
- Optimizer: AdamW with a learning rate of 2e-4 and a Cosine Scheduler.
When to Use This Model
This model is ideal for applications requiring automated analysis of Korean police reports, such as:
- Assisting law enforcement with initial case assessment.
- Automating the categorization of crime incidents.
- Generating structured summaries from unstructured police narratives.
It provides a focused solution for a niche domain, offering specialized performance where general-purpose LLMs might lack domain-specific understanding.