inclusionAI Releases SingProbe Streaming Guardrail for DeepSeek-V4.1-Flash
Start here
inclusionAI released SingProbe, a streaming guardrail probe for DeepSeek-V4.1-Flash-1017 that reads three layers of the base model's hidden states during generation and scores query intent, response unsafety and hallucination risk at each token. The 10.10M-parameter probe taps layers 12, 25 and 39 with 8 attention heads, a head dimension of 64 and a 2,048-token window; its bfloat16 weights are public. inclusionAI reports F1 of 0.8756 on query intent and 0.8668 on response safety, with streaming AUC of 0.9907 and hallucination AUC of 0.7951. Release thresholds are calibrated at 0.5 for response unsafety and 0.5 for hallucination. Running it requires the token-probe build from SGLang branch 261029, the wheel installed, both model directories mounted and eight H20 GPUs.
Generated from reports · updated 7 h ago
Report timeline
Follow the reports to see the event from different sides.
- inclusionAI ModelsDeepSeek-V4.1-Flash-1017-singprobe: SingProbe streaming guardrail released
SingProbe is an intrinsic streaming guardrail for DeepSeek-V4.1-Flash-1017 that reads three layers of the base model's hidden states during generation and scores query intent, response unsafety and hallucination risk at each token. The 10.10M-parameter probe taps layers 12, 25 and 39 with 8 attention heads, a head dimension of 64 and a 2,048-token window, and its bfloat16 weights are released. It scores F1 0.8756 on query intent and 0.8668 on response safety, with streaming AUC of 0.9907 and hallucination AUC of 0.7951.
Heat for this event
- Current, comparable range
- 6
- Peak, comparable range
- 8November 19 20:00
- Change in 24 hours
- –
The trend compares the same participants over hours observed continuously, so its range can be narrower than the current heat count. Hover or tap the chart for hourly heat; the arrow keys move through it.