Feng613/SleepVLM-3B

VISIONPricing:Input $0.32 / Cached $0.016 / Output $1.6Concurrent Unit Cost:1Model Size:3BQuant:BF16Context Size:32kPublished:Mar 17, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Feng613/SleepVLM-3B is a 3.1 billion parameter vision-language model developed by Guifeng Deng and colleagues, fine-tuned from Qwen2.5-VL-3B-Instruct. It specializes in explainable automated sleep staging from polysomnography (PSG) recordings. The model generates clinician-readable natural-language rationales, citing specific AASM scoring rules for each 30-second epoch, making its sleep staging decisions auditable. It processes rendered multi-channel PSG waveform images to predict sleep stages and provide rule-grounded explanations.

Loading preview...

SleepVLM-3B: Explainable Sleep Staging

SleepVLM-3B is a specialized 3.1 billion parameter vision-language model designed for explainable automated sleep staging from polysomnography (PSG) recordings. Developed by Guifeng Deng and colleagues, this model addresses a critical need for transparency in medical AI by providing not just a sleep stage prediction, but also a natural-language rationale grounded in clinical scoring rules.

Key Capabilities & Features

  • Rule-Grounded Explanations: Unlike traditional black-box classifiers, SleepVLM-3B generates detailed, clinician-readable rationales for its sleep staging decisions, citing specific AASM scoring rules for every 30-second epoch.
  • Vision-Language Integration: It processes rendered multi-channel PSG waveform images (three consecutive 30-second epochs) as input.
  • Output: Produces a predicted sleep stage (W/N1/N2/N3/R), applicable AASM rule identifiers, and a structured natural-language explanation.
  • Fine-tuned Architecture: Built upon Qwen2.5-VL-3B-Instruct, SleepVLM-3B underwent a two-phase training pipeline: waveform-perceptual pre-training (WPT) and rule-grounded supervised fine-tuning (SFT) using expert annotations from the MASS-EX dataset.

Intended Use

SleepVLM-3B is primarily intended for research purposes in explainable automated sleep staging. It is suitable for sleep medicine researchers, clinical informatics researchers, and AI/ML researchers focusing on interpretable medical AI. It is crucial to note that this model is not validated for clinical diagnostic use and should not replace professional sleep technologist scoring.

Finetunes
2 models