KGESH/nsfw-bge-m3-v5

NSFW BGE-M3 v5 (Category-Aware)

NSFW keyword embedding model with category-level discrimination.

Improvements over v4

• —Co-occurrence based tag clustering from 23M+ adult videos (XVideos + Big-Porn)
• —Hard negatives: different NSFW categories (creampie vs deepthroat)
• —8557 clusters from 2.1M unique tags
• —Multilingual pairs (en/ko/ja) from Danbooru aliases

Training Data

• —172K training pairs (138K positive, 1.7K hard negative, 1.5K easy negative, 31K multilingual)
• —Stage 1: MNR Loss, 2 epochs, lr=1e-5
• —Stage 2: OnlineContrastiveLoss, 5 epochs, lr=5e-6, margin=0.85