Manhph2211/PulseLM

PulseLM: A Foundation Dataset and Benchmark for PPG-Text Learning

Hung Manh Pham*   Jinyang Wu*   Xiao Ma   Yiming Zhang   Yixin Xu   Aaqib Saeed   Bin Zhu†   Zhou Pan†   Dong Ma†

* Equal contribution    † Corresponding authors

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
import neurokit2 as nk

model = AutoModelForCausalLM.from_pretrained(
    "Manhph2211/PulseLM",
    trust_remote_code=True,
    dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Manhph2211/PulseLM", trust_remote_code=True)
model.eval()

device = next(model.parameters()).device
ppg = nk.ppg_simulate(duration=10, sampling_rate=125, heart_rate=50)
ppg = torch.from_numpy(ppg).unsqueeze(0).to(device) 

messages = [
    {"role": "system", "content": (
        "You are a physiological signal analysis expert. "
    )},
    {"role": "user", "content": (
        "Classify the heart rate level shown in this PPG segment."
    )},
]


text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(device)

with torch.no_grad():
    output_ids = model.generate(
        input_ids=inputs["input_ids"],
        attention_mask=inputs["attention_mask"],
        ppg=ppg,
        max_new_tokens=32,
        do_sample=False,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.pad_token_id,
    )

print(tokenizer.decode(output_ids[0], skip_special_tokens=True).strip())

Citation

If you find this work useful, please consider citing our paper:

@misc{pham2026pulselmfoundationdatasetbenchmark,
      title={PulseLM: A Foundation Dataset and Benchmark for PPG-Text Learning}, 
      author={Hung Manh Pham and Jinyang Wu and Xiao Ma and Yiming Zhang and Yixin Xu and Aaqib Saeed and Bin Zhu and Zhou Pan and Dong Ma},
      year={2026},
      eprint={2603.03331},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.03331}, 

}