inclusionAI/LingSante-Pulse-1.0
inclusionAI/LingSante-Pulse-1.0
inclusionAI released the open-source physiological signal encoder model LingSante-Pulse-1.0, which uses the MIT license and requires a separate inference package to be installed. The model supports input [T], [C,T], [B,C,T], or aligned channel dictionaries and returns CLS and PatchMean features in the shape [B,C*256], with channels concatenated in input order. Fixed channel vectors use BAAI/bge-small-en-v1.5 for CLS pooling and L2 normalization.
Physiological signal encoder, released under the MIT license. Install the separate inference-code package before using these model assets.
Usage
The model repository is inclusionAI/LingSante-Pulse-1.0. After publication, this complete example downloads the model on first use:
import numpy as np
from lingsante_pulse import LingSantePulse
model = LingSantePulse.from_pretrained()
x = np.random.default_rng(42).normal(size=(2, 1, 1280)).astype(np.float32)
features = model.encode(x, sample_rate=128, channels=[{"modality": "ECG", "lead": "II"}])
print(features["cls"].shape) # [2, 256]
print(features["patch_mean"].shape) # [2, 256]
To load downloaded assets from the current model directory instead, use
LingSantePulse.from_pretrained(".", local_files_only=True).
revision pins a commit and local_files_only=True uses local/cached assets.
Files and interface
model.safetensors: encoder parameters.config.json: model format and automatic file-integrity checks.channels.npz: fixed channel-condition vectors and names.channels.json: physical descriptions of the channel identities.
Inputs use [T], [C,T], [B,C,T], or aligned channel dicts. Specify the actual
sampling rate and physical channel identities. The code package handles fixed
128-Hz preprocessing, observed samples, and storage-padding masks.
CLS and observed-valid PatchMean are returned separately as [B,C*256], with
channels concatenated in input order. return_tokens=True also returns ordered
patch features and validity masks at 0.25-s intervals. These are signal
representations; task-specific prediction heads are not included.
Channel-vector provenance
The fixed vectors were generated with BAAI/bge-small-en-v1.5, using CLS pooling and L2 normalization. Only the resulting vectors are included; no upstream encoder code, weights, or tokenizer is distributed here.
Source: Ant Group inclusionAI · Hugging Face new models · huggingface.co