AZET Nunchi 1
AZET’s first own model. Give it a Korean text message, messenger chat or call transcript; it answers whether it is a scam, which kind, why, and what the reader should do.
azet/nunchi-1 on this APITry it
curl https://api.azet.io/v1/chat/completions -H "Authorization: Bearer $AZET_API_KEY" -H "Content-Type: application/json" \
-d '{"model": "azet/nunchi-1", "messages": [{"role": "user", "content": "[CJ대한통운] 주소 불일치로 보관중입니다. 주소 수정 http://cj-kr.top/a1"}]}'
The instructions are built in: send only the message. Answers are AI-generated judgments, not a guarantee — when money or identity is involved, check with the organisation through its official number.
How it was built
Base model: Qwen2.5-Coder-32B-Instruct (Apache-2.0). AZET trained a LoRA adapter (rank 8, attention query and value projections) on 2,030 examples: Korean messages across 11 scam types and 8 kinds of normal messages, written for training by the open-weight model Qwen3-30B-A3B (Apache-2.0), and public voice-phishing call transcripts (HyaDoo/ko-voicephishing-binary-classification, Apache-2.0). Normal messages were kept only when their links point to the sender’s official domain. The adapter runs on Cloudflare Workers AI.
Measured through this API
60 Korean text messages from the public jmjmjm3/kor-smishing-message dataset (30 smishing, 30 normal; CC BY-NC-SA, used for evaluation only, never for training), measured 2026-10-08. Scam caught = share of scam messages flagged scam or suspicious. False alarms = share of normal messages flagged.
| System | Scam caught | False alarms | Note |
|---|---|---|---|
| Sweep phishing blocklist (200,000 hosts, URL match) | 0.0% | 0.0% | measured on 600 items of the same dataset (300 smishing, 300 normal) |
| Qwen2.5-Coder-32B-Instruct, same instructions | 90.0% | 16.7% | 27/30 caught, 5/30 false alarms |
| AZET Nunchi 1 | 86.7% | 3.3% | 26/30 caught, 1/30 false alarms |
Larger runs on our Mac
Same model and adapter, 4-bit build, measured locally 2026-10-08. Nunchi 1 trades some recall for far fewer false alarms: the base model flags most messages it sees.
| Test set | n | Base: caught / false alarms | Nunchi 1: caught / false alarms |
|---|---|---|---|
| Public smishing dataset (evaluation only) | 300 | 96.0% / 20.0% | 88.7% / 6.7% |
| Real inbox of a team member’s phone (613 normal, 8 gambling spam; kept on our machine) | 621 | 87.5% / 64.8% | 75.0% / 7.5% |
| Public voice-phishing call transcripts (held out) | 100 | 97.6% / 39.0% | 90.2% / 6.8% |
| Generated messages (held out) | 195 | 99.2% / 41.0% | 89.7% / 5.1% |
60 messages through the API is a small sample; we will publish larger API runs as they are measured. Most remaining false alarms are short bank and card notifications, which the next version targets. Known limits: Korean only; the training messages are synthetic; capacity is limited during the beta (requests may return 429 late in the UTC day).