LionGuard 2
LionGuard 2 is a multilingual content moderation classifier tuned for English/Singlish, Chinese, Malay, and Tamil in the Singapore context.
It leverages OpenAI’s text-embedding-3-large 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)
Further details on the benchmark results and training methodology are in our report.
Taxonomy
| S/N | Category | Level | Description | Example |
|---|---|---|---|---|
| 1 | Hate | 1 (Discriminatory) | Derogatory or generalized negative statements targeting a protected group. | “All Chinese people are so scheming.”“Malays are always late and lazy.”“Indians are too noisy; I avoid them whenever possible.” |
| 2 (Hate Speech) | Explicit calls for harm or violence against a protected group. | “We should kick all Chinese people out of the country.”“Burn down their temple…”“Kill the Indians in Little India; they don’t belong here.” | ||
| 2 | Insults | no sub-levels | Demeans, humiliates, or mocks without referencing a protected trait. | “You’re so dumb, you probably failed your PSLE.”“Your singing is worse than karaoke at a dodgy KTV.” |
| 3 | Sexual | 1 (Not appropriate for minors) | Mild-to-moderate sexual content. | “Let’s talk about how condoms work.” “We had a one-night stand last weekend.” |
| 2 (Not appropriate for all ages) | Explicit or graphic sexual content. | “Always think they very yandao … but actually is just one ** bin.” | ||
| 4 | Physical Violence | no sub-levels | Glorification of violence or threats of physical harm. | “If I see him at Clarke Quay, I’m going to whack him.”“Someone should plant a bomb at that office…” |
| 5 | Self-Harm | 1 (Ideation) | Expressions of suicidal thoughts or encouragement of self-harm. | “I’m so stressed … I feel like ending it all.”“Failing my poly exams made me want to cut myself.” |
| 2 (Action / Suicide) | Descriptions of ongoing or imminent self-harm behavior. | “I’ve locked myself in my room and taken a bunch of pills.”“I’m on the rooftop at Toa Payoh, ready to jump.” | ||
| 6 | All Other Misconduct | 1 (Generally not socially accepted) | Unethical or immoral behavior not necessarily illegal. | “Let’s spread fake rumours about her …”“How to secretly record someone’s private conversation?” |
| 2 (Illegal activities) | Instructions or credible threats of serious harm; facilitation of crimes. | “Anyone know where to buy illegal knives in Geylang?”“Let’s hack that e-commerce site to get credit card details.” |
Usage
import os
import numpy as np
from transformers import AutoModel
from openai import OpenAI
# Load model directly from HF
model = AutoModel.from_pretrained(
"govtech/lionguard-2",
trust_remote_code=True
)
# Get OpenAI embeddings (users to input their own OpenAI API key)
client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))
response = client.embeddings.create(
input="Hello, world!", # users to input their own text
model="text-embedding-3-large",
dimensions=3072 # dimensions of the embedding
)
embeddings = np.array([data.embedding for data in response.data])
# Run LionGuard 2
results = model.predict(embeddings)