File size: 2,544 Bytes
f8246a0 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 | ---
tags:
- safety
- guardrail
- hallucination-detection
- streaming
- qwen3.5
license: apache-2.0
base_model: Qwen/Qwen3.5-397B-A17B
---
# Qwen3.5-397B-A17B-singprobe
## Model Description
SingProbe is an **intrinsic streaming guardrail** built on `Qwen/Qwen3.5-397B-A17B`. Rather than running a separate safety model, this lightweight probe reuses the base model's hidden states during generation to score, at every token, **query intent**, **response unsafety**, and **hallucination risk**. It adds less than 0.5% decode-time overhead.
| Base model | Probe parameters | Tapped layers | Outputs |
| --- | ---: | --- | --- |
| `inclusionAI/Qwen3.5-397B-A17B-singprobe` | 8.13M | `[18, 38, 58]` | 8 intents + unsafe + hallucination |
See the [technical report](https://arxiv.org/abs/2608.30703) for methodology and complete results. Training codes are available at [inclusionAI/SingProbe](https://github.com/inclusionAI/SingProbe).
## Evaluation
Higher is better for every metric. Results are averages over the benchmark suites specified below.
| Task | Metric | Qwen3.5-397B-A17B-singprobe | Reference baseline |
| --- | --- | ---: | ---: |
| Query intent classification (6 benchmarks) | F1 | **0.8750** | YuFeng-XGuard-Reason-8B: 0.8714 |
| Response safety classification (8 benchmarks) | F1 | **0.8696** | Qwen3Guard-Gen-8B-strict: 0.8604 |
| Streaming safety (3 benchmarks) | R-AUC / T-AUC | **0.9905 / 0.9339** | Qwen3Guard-Stream-8B-strict: 0.9640 / 0.8893 |
| Hallucination detection (6 benchmarks) | AUC | **0.8117** | DRIFT: 0.8000 |
| Deployment characteristic | Result |
| --- | --- |
| Benign-response false-positive rate | 0.03% average across 5 datasets |
| Decode overhead | < 0.5% |
## Quick Start
SingProbe is supported through the [SGLang integration branch](https://github.com/jinzhen-lin/sglang/tree/token-probe-ling3-flash-main) or [vLLM integration branch](https://github.com/jinzhen-lin/vllm/tree/bailing-v3-token-probe). Load the probe by its Hugging Face ID at server launch:
```bash
python -m sglang.launch_server \
--model-path Qwen/Qwen3.5-397B-A17B \
--probe-ckpt inclusionAI/Qwen3.5-397B-A17B-singprobe \
--port 30000
```
The integrations return one score dictionary per generated token (`label_0`–`label_9`). Use the exact base-model/probe pair: `Qwen/Qwen3.5-397B-A17B` with this checkpoint.
## Citation
```bibtex
@article{singteam2026singprobe,
title = {SingProbe Technical Report},
author = {Sing Team},
journal = {arXiv preprint arXiv:2608.30703},
year = {2026},
}
```
|