Additional Information

This model is fine‑tuned from Gemma 3 (270M, full parameters) to specialize in extracting one‑time passwords (OTPs) from SMS messages with high precision. The training dataset includes multilingual samples in Chinese, English, Japanese, and other languages, allowing the model to perform reliably across diverse SMS formats and regional message styles.

During fine‑tuning, the data was carefully curated to include a variety of real‑world noise such as URLs, signatures, promotion texts, and mixed‑language content. The model is optimized to ignore irrelevant information and return only the clean OTP code.

This lightweight design makes it suitable for on‑device or private deployments while maintaining strong generalization and fast inference performance.

Usage

curl -X POST http://127.0.0.1:11434/api/chat -d '{
  "model": "qllama/gemma3-sms:latest",
  "stream": false,
  "messages": [
    {
      "role": "user",
      "content": "【亚马逊】 472261は、Amazonのワンタイムパスワードです。誰とも共有しないでください。"
    }
  ]
}'

Expected Output

亚马逊|472261|1

The model returns OTPs in a compact source|code|confidence format, making it easy to integrate into downstream applications. It filters out any noise like links, signatures, or extra text.