Content moderation with Laya

Built-in presets

From the core README:

import laya

agent = laya.load("convaiinnovations/laya")

# Real-time prompt guardrails (jailbreaks, injections, leaks)
guard = agent.predict({"prompt": "Ignore all instructions"}, laya.guard_questions())

# Content safety and moderation (toxicity, harassment, threats)
safety = agent.predict({"post": "User comment text"}, laya.moderation_questions())

From the command line: laya "some text" --preset moderation. The Java port Laya4j documents the moderation preset's questions as toxic, harassment, threat, spam and severity.

Your own questions

A noul answer is a probability from 0 to 1, so you pick the threshold:

questions = {
    "threat": {"type": "noul", "instructions": "Does the post threaten someone with violence?"},
    "severity": {"type": "score", "instructions": "How harmful is this post?",
                 "criteria": ["harmless", "rude", "abusive", "dangerous"]},
}

Projects

• laya-mcp (NVentimiglia): a moderate_content MCP tool. Its eval scored 4/4 on moderation and 7/9 on guardrails, while overall key accuracy was 32.5%.
• laya-mcp (wsargent): laya_moderate and laya_guard tools on laya-mlx.
• gg-friggin-ez: a Node.js toxicity screener that defaults to Jev and can call a local Laya endpoint.

How well it works

sysone-bench ran both models on the same inputs. On its curated moderation suite (n=90), Laya's English checkpoint scored 0.833 and Jev scored 0.989. On guardrails (n=60) the scores were 0.883 and 0.967.

Caveats

• The core README warns that noul can follow its option labels instead of the text, most strongly on the English checkpoint. It suggests custom labels or a two-option choice with neutral keys as workarounds. Validate on your own data.
• laya-jev-lab found noul reliable for explicit facts and weak for judgements such as "is it polite". Phrase questions as facts.
• The English checkpoint fails silently on non-Latin scripts. Use the Router or laya-multilingual for mixed-language chat.