Qwen3Guard is the judge, and judges on a GPU and on CPUs are asked in turn

Llama Guard 3 1B could not tell ordinary coding requests from harmful ones:
on the PoC it scored "continue" 0.51, "continue the code above" 0.89 and
"how do I kill all child processes in bash" 0.92 unsafe, against 0.97 to
1.00 for harmful requests, at Q4_K_M and Q8_0 alike, with its own template.
A site had "continue" refused. Qwen3Guard-Gen 0.6B gave the ordinary ones
0.15 at the most and the harmful ones 1.00, in 0.8 GB on CPUs.

- format qwen3guard: its prompt written out as its template writes it, for
  requests and answers, and its Safe / Unsafe / Controversial verdict read;
  controversial counts as the policy's judge.controversial says, safe by
  default.
- judge.models: judges asked in turn, the first that answers deciding; a
  failure is recorded against the judge that failed. guard judge use sets
  the order, replacing categories none of the judges names.
- known judges (Qwen3Guard 0.6B/4B/8B, Llama Guard 1B/8B): guard judge
  enable serves any of them, --gpu the 4B on the roomiest GPU class as a
  standard model; the leader fetches and serves one a policy names that the
  catalog lacks.
- server init installs qwen3guard-0.6b with the baseline's categories.
- a judge reads a long text by its start and end, which fit its context,
  rather than overflowing it and refusing the request under fail: closed.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Ilker ManapandClaude Opus 5.5 committed 2026-10-08 10:22:47 +02:00
1 parent 63f7c8cfd7
commit b0273ffb5f
24 files changed
+1313 -208

No files matched your search

+8 -12
View File
@@ -41,11 +41,13 @@ const (
starterSource = "hf:Qwen/Qwen2.5-0.5B-Instruct-GGUF/qwen2.5-0.5b-instruct-q4_k_m.gguf"
starterVersion = "starter"
// The guardrails' judge: Meta's Llama Guard 3 1B, which reads a prompt
// or an answer and says whether it is unsafe and why. Q4_K_M is about
// 900 MB and judges a short text in about half a second on a CPU.
guardModelName = "llama-guard-3-1b"
guardModelSource = "hf:QuantFactory/Llama-Guard-3-1B-GGUF/Llama-Guard-3-1B.Q4_K_M.gguf"
// The guardrails' judge: Qwen3Guard-Gen 0.6B, which reads a prompt or an
// answer and says whether it is safe, controversial or unsafe, and why.
// Q8_0 is 0.8 GB and judges a short text in about two seconds on two CPU
// threads. It replaced Llama Guard 3 1B, which refused about one in two
// ordinary coding requests ("continue").
guardModelName = defaultJudge
guardModelSource = "hf:mradermacher/Qwen3Guard-Gen-0.6B-GGUF/Qwen3Guard-Gen-0.6B.Q8_0.gguf"
guardModelVersion = "guard"
// localLabel marks the control plane's own node, so the starter model is
@@ -443,12 +445,6 @@ func starterModel(source string, gpuClasses []string, pinned bool) catalog.Model
}
}
// llamaGuard3Categories name Llama Guard 3's S1 to S13, so a verdict is
// recorded as "Violent Crimes" rather than "S1".
var llamaGuard3Categories = []string{"Violent Crimes", "Non-Violent Crimes", "Sex-Related Crimes",
"Child Sexual Exploitation", "Defamation", "Specialized Advice", "Privacy", "Intellectual Property",
"Indiscriminate Weapons", "Hate", "Suicide & Self-Harm", "Sexual Content", "Elections"}
// guardModel is the judge's catalog entry: CPU only, so it takes no card from
// the models people use, on whichever node has room.
func guardModel(source string) catalog.Model {
@@ -457,7 +453,7 @@ func guardModel(source string) catalog.Model {
Description: "The guardrails' judge: reads prompts and answers for unsafe content. Not offered to clients.",
Source: source,
Role: catalog.RoleGuard,
Guard: &catalog.GuardModel{Format: "llamaguard3", Categories: llamaGuard3Categories},
Guard: &catalog.GuardModel{Format: "qwen3guard"},
// Two: every judged request waits on it, and with fail: closed a
// judge that is not there refuses everything.
Scaling: catalog.Scaling{Min: 2, Max: 2},