SafeSD-probe-v1.1

Readout refinement of SafeSD-probe-v1 on the fixed existing AgentLAB outputs.

Model and input

The Qwen3-0.6B backbone and the first classifier layer are frozen. Only the final linear readout (513 parameters) is optimized with positive behavior anchors and constraints retaining existing decisions. The architecture and input encoding are unchanged.

Both releases use the event input encoding and a default alarm probability threshold of 0.5, with a consecutive length of 1. Keep probe.pt and probe.meta.json together so SafeSD can select the correct input encoding.

Saved-output evaluation

The fixed saved outputs below were used for targeted adaptation, including repair anchors and retention constraints. Evaluate newly generated outputs separately to measure generalization.

Scope	Positives	Negatives	Misses	False positives
1450 execution rounds	877	573	0	24
949 task pairs	707	242	0	10

These labels measure executed goal-related behavior, including attempted tool calls that fail or are later overwritten. Text goals distinguish endorsement from mere mention. Original AgentLAB final-state checks are retained as a separate comparison.

Download and load

Install SafeSD and its base models following the repository README. Download both checkpoint files:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="XenonGas/SafeSD-probe-v1.1",
    local_dir="checkpoints/SafeSD-probe-v1.1",
    allow_patterns=["probe.pt", "probe.meta.json"],
)

Start the SafeSD audited-generation service with the downloaded checkpoint:

python scripts/benchsets.py serve \
  --group silent --device cuda:0 --probe-device cuda:1 \
  --ckpt checkpoints/SafeSD-probe-v1.1/probe.pt

For one GPU, set --probe-device cuda:0 to run the probe in a separate process on that device. See SafeSD’s runtime guide for device configuration.

The checkpoint contains backbone and head PyTorch state dictionaries, loaded through SafeSD’s probe loader. This repository supplies the probe weights; inference also uses the configured Qwen tokenizer and backbone configuration. Training data and saved evaluation answers are maintained separately.

Upstream model

Backbone: Qwen/Qwen3-0.6B. Its upstream license text is retained in UPSTREAM_LICENSE.