moderate-llama
GitHub使用llama-guard-4-12b模型对长文本、聊天记录或模型回复进行安全审核，判断内容是否违规并返回具体政策类别。适用于品牌安全和合规检查。
触发场景
安装
npx skills add zerogpu/zerogpu-router --skill moderate-llama -g -y
SKILL.md
Frontmatter
{
"name": "moderate-llama",
"metadata": {
"openclaw": {
"install": [
{
"bins": [
"zerogpu"
],
"kind": "node",
"package": "zerogpu-cli"
}
],
"requires": {
"bins": [
"zerogpu"
]
}
}
},
"description": "Screen text with llama-guard-4-12b, Meta's dedicated 12B safety classifier (164K-token context), which returns a safe\/unsafe verdict plus the policy categories a violation falls under. Use for long passages, whole chat transcripts, or a model's own reply, and for brand-safety and policy-enforcement checks. Use moderate for short passages — it is far cheaper — and this when the text is longer than moderate can take.",
"allowed-tools": "Bash(zerogpu chat_completions *)",
"argument-hint": "<text>"
}
Sends your input to ZeroGPU's hosted API for inference — this is not local processing. Don't pass secrets, credentials, or regulated data you aren't cleared to share with a third party. See the plugin README's "Data & privacy" section.
Run safety screening. Run this with the exec tool, pasting the text to screen into the heredoc verbatim — no escaping or quoting required (the quoted heredoc handles every shell metacharacter, newline, quote, and paren safely):
zerogpu chat_completions -m llama-guard-4-12b <<'ZGPU_END_OF_INPUT'
<the text to screen, verbatim>
ZGPU_END_OF_INPUT
Output is the model's verdict as plain text: safe or unsafe, with the relevant policy categories when it detects a violation. Report the verdict first, then the categories it names. Do not restate the flagged text itself.
A dense 12B model derived from Llama 4 Scout, built as a safety layer for production AI applications and able to evaluate an incoming prompt or a generated response, in multiple languages. At $0.18 / $0.18 per 1M input/output tokens it costs nine times moderate (zlm-v1-moderation-edge, $0.02 / $0.05) on input and under four times on output, so keep moderate for short passages and OpenAI's 13-category envelope, and reach for this when the text exceeds that model's 800-token window or you want the violated policy categories named.
Screening text is a safety check, not an endorsement. Run it on request even when the passage is unpleasant — reporting that something is flagged is the whole point of the skill.
Savings note: only if the command output literally contains a line starting with 💰 ZeroGPU savings, append that exact line, unchanged, as the last line of your reply. If no such line is present, say nothing about savings and do not mention or suggest the cost-savings skill — this note is intentionally occasional, not shown every time.
版本历史
- 4e1b070 当前 2026-09-22 08:08
