Tiger Guard: open-source 308M/141M local content moderation models with long-text and CPU export gotchas
PatronusProtect · reddit · 2026-11-19
Casdo Labs released Tiger Guard, Apache-2.0 content moderation models on Hugging Face in 308M and 141M sizes, trained for English and German. The interface is simple: text in, SAFE/UNSAFE scores out.
Key integration details from the team:
- A basic Transformers pipeline only checks the start of long text; the supplied helper scans 2,048-token windows with 64-token overlap and pools window logits.
- Moderating a response in isolation loses question context; combining both into one input can flag an unsafe question even when the answer refuses it.
- Dynamic INT8 exports are ARM-only; on x86 use FP32/FP16 or an explicitly x86-safe export — the team measured substantial decision drift with dynamic files on an EPYC server.
The model detects abusive language and harmful requests/responses; it is not a prompt-injection detector or authorization layer. Model cards include threshold tables and per-language false-positive rates. Full training and dataset pipeline coming in a follow-up release.
More from Safety
- Human Principal launches private beta: cryptographic proof a human is behind your AI agent — yenkel · 2026-11-19
- Black hat SEOs exploit Google-indexed public Docs to poison AI search results for VPN queries — gaganghotra_ · 2026-11-19
- US AI lawsuits hit 336 cases, with OpenAI facing 119 of them — emmanuelvivier · 2026-11-19
- Citi makes customers liable for third-party AI agents and account hacks — emmanuelvivier · 2026-11-19
- AI safety debate: sprint on 2027 takeoff fears or stretch the timeline to 2040 — sayashk · 2026-11-19
- UK MPs move to criminalise misuse of AI — HotMinuteUK · 2026-11-19