inclusionAI releases streaming guardrail probe for Ling-3.1-flash-1019
Start here
inclusionAI released Ling-3.1-flash-1019-singprobe, a streaming guardrail probe that attaches to the Ling-3.1-flash-1019 model and scores query intent, response unsafety and hallucination risk at every token. The probe reads hidden states at layers 23, 47 and 71 and has 8.13M parameters; the released weights use bfloat16, with calibrated thresholds of 0.5 for response unsafety and 0.5 for hallucination. inclusionAI reports F1 of 0.8687 on query intent and 0.8707 on response safety, plus 0.9844 response AUC and 0.9411 token AUC on streaming safety, with a 0.008% false-positive rate on benign responses. These are the publisher's own figures and have not been independently verified.
Generated from reports · updated 7 h ago
Report timeline
Follow the reports to see the event from different sides.
- inclusionAI ModelsLing-3.1-flash-1019-singprobe: A Streaming Guardrail Probe for Ling-3.1-flash-1019
Ling-3.1-flash-1019-singprobe is an intrinsic streaming guardrail for Ling-3.1-flash-1019 that taps hidden states at layers 23, 47 and 71 with an 8.13M-parameter probe to score query intent, response unsafety and hallucination risk at every token. It reaches F1 0.8687 on query intent and 0.8707 on response safety, plus 0.9844 response AUC and 0.9411 token AUC on streaming safety, with a 0.008% benign-response false-positive rate.
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.