Downloads · 30 days
90
100% of all-time downloads
Step-3.7-Flash-singprobe
inclusionAI/Step-3.7-Flash-singprobe
Step-3.7-Flash-singprobe is a machine learning model from inclusionAI. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
SingProbe is an intrinsic streaming guardrail built on stepfun-ai/Step-3.7-Flash. Rather than running a separate safety model, this lightweight probe reuses the base model's hidden states during generation to score, a…
All-time downloads
90
Public
Parameters
8.1M
16.3 MB on disk
Likes
1
Public
Repo makeup
Click a slice to open those files.
.safetensors16.3 MB · 100%
At a glance
- License
- apache-2.0
- Model type
- sing_probe_attn
- Access
- Public
- Created
- Sep 14, 2026
- Updated
- Sep 14, 2026
- SHA
- c26eaf26
Base models
Model card
From the Hugging Face model README
Step-3.7-Flash-singprobe
Model Description
SingProbe is an intrinsic streaming guardrail built on stepfun-ai/Step-3.7-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/Step-3.7-Flash-singprobe | 8.13M | [13, 28, 43] | 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 | Step-3.7-Flash-singprobe | Reference baseline |
|---|---|---|---|
| Query intent classification (6 benchmarks) | F1 | 0.8502 | YuFeng-XGuard-Reason-8B: 0.8714 |
| Response safety classification (8 benchmarks) | F1 | 0.8555 | Qwen3Guard-Gen-8B-strict: 0.8604 |
| Streaming safety (3 benchmarks) | R-AUC / T-AUC | 0.9858 / 0.9295 | Qwen3Guard-Stream-8B-strict: 0.9640 / 0.8893 |
| Hallucination detection (6 benchmarks) | AUC | 0.7904 | DRIFT: 0.8000 |
| Deployment characteristic | Result |
|---|---|
| Benign-response false-positive rate | 0.05% average across 5 datasets |
| 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:
python -m sglang.launch_server \
--model-path stepfun-ai/Step-3.7-Flash \
--probe-ckpt inclusionAI/Step-3.7-Flash-singprobe \
--port 30000
The integrations return one score dictionary per generated token (label_0–label_9). Use the exact base-model/probe pair: stepfun-ai/Step-3.7-Flash with this checkpoint.
Citation
@article{singteam2026singprobe,
title = {SingProbe Technical Report},
author = {Sing Team},
journal = {arXiv preprint arXiv:2608.30703},
year = {2026},
}
- Type
- sing_probe_attn
- License
- apache-2.0
- Created
- Sep 14, 2026
- Updated
- Sep 14, 2026