wikiepeidia/vnphish-dataset

VNPhish: Vietnamese financial phishing messages

2,097 Vietnamese messages labelled into four classes, built for a bachelor internship project at the University of Science and Technology of Hanoi (USTH), 2026: Design and Development of a Localized LLM for Vietnamese Financial Phishing Detection (Phạm Thế Minh). Code: https://github.com/wikiepeidia/Internship-project

The messages are synthetic. They were written by language models from public scam warnings, not collected from real victims.

Splits

Split	Bank impersonation	Benign	Task scam	Zalo social engineering	Total
train	595	517	306	240	1,658
validation	76	72	49	22	219
test	70	66	49	35	220

All messages built from the same source scenario share a seed_id and are in the same split, so near-duplicate rewordings never cross between train and test.

Fields

Field	Meaning
text	the message
label	bank_impersonation, zalo_social_engineering, task_scam, or benign
risk_tier	benign, suspicious, or high-risk (severity, separate from the label)
suspicious_spans	up to 3 exact substrings of text that support the label
xai_explanation	one-sentence Vietnamese explanation
source	which generator wrote the message (see below)
seed_id	the source scenario the message was written from

How it was built

1. Public scam-warning articles were collected from tinnhiemmang.vn. seed_sources.csv lists each seed's ID, URL and collection time. The article text itself is not redistributed.
2. Messages were generated one class at a time. The prompt names the class, so the label is set when the message is written.
   • synthetic_claude (745 messages): Claude 3.5 Haiku.
   • synthetic_openai_compatible (1,352 messages): GPT-5.6 Sol, reached through an OpenAI-compatible API, including the rewritten Zalo class.
3. Every message was scored by a separate judge model on five 1–5 checks (realism, label correctness, Vietnamese–English mixing, risk tier, span accuracy). A message passes only if all five are at least 3. 1,395 of 2,097 pass (66.5%); failing messages are still included. Per-message scores are in judge_scores.jsonl (join on split + row_index, the line number within the split file, starting at 0).
4. A reviewer checked 100 messages by hand (44 passed; agreed with the judge on
   87. and reviewed all 324 messages the judge flagged as possibly mislabelled.
5. 60 of the 147 seed_id values come from rewritten Zalo scenarios rather than collected articles, so they have no entry in seed_sources.csv.

Changes for publication

Phone numbers and email addresses were masked in this public copy: 666 Vietnamese mobile-number patterns in 353 messages keep their first digits and have the rest replaced with x, and 5 email addresses became user@example.com. Spans still match the text. Models in the thesis were trained and evaluated on the unmasked version.

Limitations

• Synthetic messages: they do not show how real scam traffic looks.
• Labels come from generation plus checking, not full human annotation.
• The judge model is not fully independent of the generators.
• One reviewer, and only a sample was checked by hand.

Intended use

Research and teaching on Vietnamese scam detection. Not for building systems that generate scam messages.