lionguard-2.1

LionGuard 2.1

LionGuard 2.1 is a multilingual content moderation classifier tuned for English/Singlish, Chinese, Malay, and Tamil in the Singapore context.

It leverages Gemini's gemini-embedding-001 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)

---

Taxonomy

---

Usage

import os
import numpy as np
from transformers import AutoModel
from google import genai

# Load model directly from HF
model = AutoModel.from_pretrained("govtech/lionguard-2.1", trust_remote_code=True)

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

# Get embeddings (users to input their own Gemini API key)
client = genai.Client(api_key=os.getenv("GEMINI_API_KEY"))
response = client.models.embed_content(
    model="gemini-embedding-001",
    contents=texts
)
embeddings = np.array([emb.values for emb in response.embeddings])

# Run inference
results = model.predict(embeddings)