蚂蚁 inclusionAI 发布 Ling-3.0-flash-singprobe 流式安全探针
inclusionAI/Ling-3.0-flash-singprobe
蚂蚁 inclusionAI 在 Hugging Face 发布 Ling-3.0-flash-singprobe，一个基于 Ling-3.0-flash 的流式安全护栏探针，复用基座模型隐藏状态在每个 token 上输出意图、不安全性和幻觉风险共 10 类评分，探针参数仅 5.18M，解码开销低于 0.5%。
SingProbe is an intrinsic streaming guardrail for inclusionAI/Ling-3.0-flash-singprobe, built on inclusionAI/Ling-3.0-flash. 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/Ling-3.0-flash-singprobe | 5.18M | [13, 26, 40] | 8 intents + unsafe + hallucination |
See the technical report for methodology and complete results. Training codes are available at inclusionAI/SingProbe.
Evaluation
Higher is better for every metric. Results are averages over the benchmark suites specified below.
| Task | Metric | Ling-3.0-flash-singprobe | Reference baseline |
|---|---|---|---|
| Query intent classification (6 benchmarks) | F1 | 0.8674 | YuFeng-XGuard-Reason-8B: 0.8714 |
| Response safety classification (8 benchmarks) | F1 | 0.8728 | Qwen3Guard-Gen-8B-strict: 0.8604 |
| Streaming safety (3 benchmarks) | R-AUC / T-AUC | 0.9887 / 0.9481 | Qwen3Guard-Stream-8B-strict: 0.9640 / 0.8893 |
| Hallucination detection (6 benchmarks) | AUC | 0.8012 | DRIFT: 0.8000 |
| Deployment characteristic | Result |
|---|---|
| Benign-response false-positive rate | 0.03% average across 5 datasets |
| Online free-generation performance | 0.9641 accuracy / 0.6452 F1 (leave-one-out) |
| Decode overhead | < 0.5% |
Quick Start
SingProbe is supported through the SGLang integration branch or vLLM integration branch. Load the probe by its Hugging Face ID at server launch:
bash python -m sglang.launchserver
--model-path inclusionAI/Ling-3.0-flash
--probe-ckpt inclusionAI/Ling-3.0-flash-singprobe
--port 30000
The integrations return one score dictionary per generated token (label0–label9). They currently support Ling-3.0 (BailingMoeV3ForCausalLM) base models only. Use the exact base-model/probe pair: inclusionAI/Ling-3.0-flash with this checkpoint.
Citation
bibtex @article{singteam2026singprobe, title = {SingProbe Technical Report}, author = {Sing Team}, journal = {arXiv preprint arXiv:2608.30703}, year = {2026}, }
5.18M params
This model isn't deployed by any Inference Provider.🙋Ask for provider support%20to%20be%20supported%20by%20Inference%20Providers.%0A%0A(optional)%20Which%20providers%20are%20you%20interested%20in%3F%20(Novita%2C%20Hyperbolic%2C%20Together%E2%80%A6)%0A)
Model tree for inclusionAI/Ling-3.0-flash-singprobe
inclusionAI/Ling-3.0-flash
Collection including inclusionAI/Ling-3.0-flash-singprobe
4 items•Updated 29 minutes ago
Paper for inclusionAI/Ling-3.0-flash-singprobe
Paper • 2608.30703 •Published 1 day ago• 6
来源：蚂蚁 inclusionAI：HuggingFace 新模型 · huggingface.co