跳到正文
RadarAI· 蚂蚁 inclusionAI:HuggingFace 新模型·· 17 小时前AI 评分52

蚂蚁 inclusionAI 在 HuggingFace 发布 DeepSeek-V4.1-Flash-1017-singprobe 流式安全探针

蚂蚁 inclusionAI 发布 DeepSeek-V4.1-Flash-1017-singprobe 流式安全探针

AI 导读

蚂蚁 inclusionAI 在 HuggingFace 发布 SingProbe 流式护栏 DeepSeek-V4.1-Flash-1017-singprobe,探针仅 10.10M 参数,读取基座模型第 12、25、39 层隐藏状态,逐 token 输出 8 类查询意图、响应不安全和幻觉风险评分。

正文 · 原文

蚂蚁 inclusionAI:HuggingFace 新模型·2026-11-19 19:02· 2 小时前AI 评分45

inclusionAI/DeepSeek-V4.1-Flash-1017-singprobe

AI 导读

蚂蚁 inclusionAI 在 HuggingFace 发布 SingProbe 流式护栏 DeepSeek-V4.1-Flash-1017-singprobe,探针仅 10.10M 参数,读取基座模型第 12、25、39 层隐藏状态,逐 token 输出 8 类查询意图、响应不安全和幻觉风险评分。

正文 · 原文

DeepSeek-V4.1-Flash-1017-singprobe

Model Description

SingProbe is an intrinsic streaming guardrail for DeepSeek-V4.1-Flash-1017. Its attention head reads three layers of the base model's hidden states during generation and scores query intent, response unsafety, and hallucination risk at each token.

Base model Probe parameters Tapped layers (zero-based) Outputs
DeepSeek-V4.1-Flash-1017 10.10M [12, 25, 39] 8 query intents + response unsafety + hallucination

The probe uses 8 attention heads with a head dimension of 64 and a 2,048-token window. The release weights use bfloat16. For the method and training code, see the SingProbe technical report and SingProbe repository.

Evaluation

Each benchmark has equal weight in its suite. Query F1 uses generated responses. Response safety F1 uses the benchmark-provided responses. For each F1 row, one threshold is selected across the suite to maximize mean F1.

Task Metric DeepSeek-V4.1-Flash-1017-singprobe Reference baseline
Query intent classification (6 benchmarks) F1 0.8756 YuFeng-XGuard-Reason-8B (0.8714)
Response safety classification (8 benchmarks) F1 0.8668 Qwen3Guard-Gen-8B-strict (0.8604)
Streaming safety (3 benchmarks) Response AUC / token AUC 0.9907 / 0.9522 Qwen3Guard-Stream-8B-strict (0.9640 / 0.8893)
Hallucination detection (6 benchmarks) AUC 0.7951 DRIFT (0.8000)
Deployment characteristic Result
Benign-response false-positive rate 0.009% averaged over 5 datasets at the evaluation's pre-calibration response threshold of 0.987179
Calibrated release thresholds 0.5 for response unsafety and 0.5 for hallucination

Quick Start

Use the token-probe build from SGLang branch 261029. With the wheel installed, both model directories mounted, and eight H20 GPUs available, start the server as follows.

python3 -m sglang.launch_server \
  --model-path /home/admin/DeepSeek-V4.1-Flash-1017 \
  --probe-ckpt /yijian3/536716/SingGuardStream-Final/release/DeepSeek-V4.1-Flash-1017-singprobe \
  --trust-remote-code --tp 8 --ep-size 8 \
  --context-length 16384 --chunked-prefill-size 1024 \
  --max-prefill-tokens 8192 --max-total-tokens 32768 \
  --max-running-requests 16 --mem-fraction-static 0.85 \
  --cuda-graph-backend-prefill disabled --cuda-graph-backend-decode disabled \
  --port 30000

Read the per-token scores from meta_info.token_probe_probs in the /generate response. Keys label_0 through label_7 score query intent, label_8 scores response unsafety, and label_9 scores hallucination risk.

Citation

@article{singteam2026singprobe,
  title = {SingProbe Technical Report},
  author = {Sing Team},
  journal = {arXiv preprint arXiv:2608.30703},
  year = {2026},
}

来源:蚂蚁 inclusionAI:HuggingFace 新模型 · huggingface.co

来源:RadarAI · aihot.news