BERTweet-Guard

Real-time binary content moderation microservice.

Waking up API (~15s)... BERTweet + LoRA (r=32)

Input Content

Classification Result

Awaiting...

Normal Discourse 0.0%

Flagged (Toxic / Threat / Hate) 0.0%

Calibrated Threshold: 0.6760 Margin: --

System Architecture

▹ Model: vinai/bertweet-base fine-tuned via Low-Rank Adaptation (LoRA) on the 160k-row Jigsaw Toxicity corpus.
▹ Optimization: Weighted Focal Loss (γ=1.0) applied to focus gradient updates on hard minority examples (direct threats).
▹ Thresholding: The 0.6760 decision boundary was calibrated via a validation Macro F1 sweep to protect benign political discourse while isolating genuine hostility.