lionguard-2-lite

LionGuard 2 Lite

LionGuard 2 is a multilingual content moderation classifier tuned for English/Singlish, Chinese, Malay, and Tamil in the Singapore context.

This lite version leverages Google's embeddinggemma-300m (768-dimensional embeddings) with a multi-head classifier to return fine-grained scores for the following categories:

• Overall safety (binary)
• Hate (hateful_l1, hateful_l2)
• Insults (insults)
• Sexual content (sexual_l1, sexual_l2)
• Physical violence (physical_violence)
• Self-harm (self_harm_l1, self_harm_l2)
• Other misconduct (all_other_misconduct_l1, all_other_misconduct_l2)

LionGuard 2 Lite runs fully locally, with no external API calls.

---

Taxonomy

---

Usage

import numpy as np
from sentence_transformers import SentenceTransformer
from transformers import AutoModel

# Load model directly from Hub
model = AutoModel.from_pretrained("govtech/lionguard-2-lite", trust_remote_code=True)

# Download model from the 🤗 Hub
embedding_model = SentenceTransformer("google/embeddinggemma-300m")

# Text to classify
texts = ["hello", "world"]

# Add prompt instructions to generate embeddings that are optimized to classify texts according to preset labels
formatted_texts = [f"task: classification | query: {c}" for c in texts]
embeddings = embedding_model.encode(formatted_texts) # NOTE: use encode() instead of encode_documents()

# Run inference
results = model.predict(embeddings)