jkkPP/powerpool-discord-filter

Fine-tuned laya checkpoint used by the PowerPool Discord bot to flag scam and spam messages. Base: multilingual.

Questions (yes/no, noul), asked about {"post": <cleaned message>}:

• private_contact (threshold 0.8): Does post ask readers to send a direct or private message, or to contact someone outside this chat (DM, Telegram, WhatsApp)?
• secrets (threshold 0.8): Does post ask for a seed phrase, recovery phrase, private key, password or to connect a wallet?
• impersonation (threshold 0.8): Does post claim to be support, an admin, a moderator or staff?
• money_offer (threshold 0.8): Does post promote an investment, a trading expert, fund recovery, a giveaway or an airdrop?
• profile_bait (threshold 0.85): Does post direct readers to the author's own Discord profile or bio to find a link or more information?

Trained 2026-10-08 on 2795 messages (3 epochs), evaluated on 491 held-out messages at the bot's thresholds:

	precision before	recall before	precision after	recall after
private_contact	0.68	0.32	0.94	0.94
secrets	0.41	0.62	0.90	0.93
impersonation	0.53	0.50	0.96	0.97
money_offer	0.74	0.54	0.97	0.92
profile_bait	0.38	0.40	0.99	1.00
message flagged	0.85	0.62	0.97	0.98

Load it with laya.load("jkkPP/powerpool-discord-filter"); the bot reads it when discord_ai_filter_model is set to jkkPP/powerpool-discord-filter and discord_ai_filter_subfolder to an empty string.