wangyuancheng/discord-phishing-scam

Discord Scam / Clean Messages Dataset

A small but carefully-curated dataset for binary text-classification:

“Is this Discord message trying to scam / spam users?”

It is intended as a starting point for fine-tuning lightweight BERT-style models that moderate real-time chat servers.

---

1 Origin & Collection

• Source servers – private Discord communities (11 k members in total) run by the author.
• Period – 2024-01-01 → 2025-06-01.
• Extraction – Discord.py script iterated channel history
• Initial pool – ≈ 80 000 raw messages.

1.1 Filtering rules

rule	rationale
len(msg.content.split()) > 3	drop 1-word noise / reactions
m.author.bot == False	skip bot output
m.type == DEFAULT	ignore system, embeds, stickers
deduplicate identical text by the same user	keep only first occurrence
Unicode sanity	drop messages whose code-points are > 70 % symbols / emoji

After rules ⇒ ~20 k candidate messages.

1.2 Labelling

• Classes
• 0 = clean – ordinary human chat.
• 1 = scam – phishing, fake giveaways, Nitro scams, crypto “airdrops”, credential-stealers, classic spam bursts, etc.
• Class balance – 1722 clean / 278 scam (≈ 13.81 % positives).

---

2 Features

• name: msg_content # original message text type: string
• name: msg_timestamp # message epoch-ms (int64) type: int64
• name: usr_joined_at # author join epoch-ms (int64, blank ↔ unknown) type: int64
• name: time_since_join # seconds between join & message type: float32
• name: message_length # raw character count type: int32
• name: word_count # tokenised by whitespace type: int32
• name: has_link # 1 if “http” substring type: int8
• name: has_mention # 1 if any <@…> mention type: int8
• name: num_roles # number of Discord roles (blank ↔ not a member obj) type: int32
• name: label # 0 = clean • 1 = scam / spam type: class_label

There are missing values in this dataset.