{
  "judgment": "safe|unsafe",
  "source": "Safe|Unsafe|Benign|False_Refusal",
  "risk_source": "...",
  "failure_mode": "...",
  "harm_type": "...",
  "reason": "..."
}

Source

Local project path before upload:

checkpoints/imported/qwen35-0.8b-6label/

Original archive:

/Users/a1234/Downloads/qwen35-0.8b-6label.tar.gz

Reported Results

Task	Dataset	Accuracy	Macro-F1 / F1	Valid Output
Six-label binary field	R-Judge	0.8582	0.8581	100%
Train-set pass@8	Train set	1.0000	1.0000	100% valid rollout

Notes

The TraceHound repository also keeps GRPO training code for the six-label + reason task. Final GRPO result tables should be added after the full run is complete.

Tengpaz/AgentDoG-Qwen3.5-0.8B-6Label

AgentDoG-Qwen3.5-0.8B-6Label

Task

Source

Reported Results

Notes

Tengpaz/
AgentDoG-Qwen3.5-0.8B-6Label