kdt-2-team4-newbiz/Qwen3-1.7B

Qwen3-1.7B: Smishing Explanation Generator

This model, kdt-2-team4-newbiz/Qwen3-1.7B, is built upon the Qwen/Qwen3-1.7B base and is specialized in generating concise Korean explanations for SMS phishing (smishing) messages. It operates effectively through a few-shot prompting approach, eliminating the need for extensive fine-tuning. The model's primary function is to articulate why a given SMS message is suspected of being smishing, based on identified characteristics.

Key Capabilities

• Smishing Explanation Generation: Provides a single, under 70-character Korean sentence explaining the reason for smishing suspicion.
• Few-shot Learning: Utilizes a system prompt and six distinct few-shot examples covering various smishing detection features.
• Feature-based Analysis: Explanations are generated based on detected features such as:
  • Dangerous keywords (e.g., refund, loan)
  • Inclusion of external links
  • Inducement to external contact (e.g., WhatsApp, Telegram)
  • Presence of phone numbers
  • Content related to money
  • International SMS origin

Integration and Usage

This model is designed to be part of a larger system, specifically called upon by a classifier (e.g., kdt-2-team4-newbiz/kcelectra-smishing-classifier) only when an SMS message has already been identified as smishing. It takes the SMS content and a list of detected features as input and outputs the explanation. The model's prompt structure includes a system role instruction, output format constraints, and a series of user/assistant few-shot examples to guide its response generation.