romboai/rose-1h-nmr

ROSE-1H NMR

Spectrum-side foundation model for ¹H NMR (7.8M parameters). Shared encoder, five heads: denoise, pair, retrieval, forward, peak.

Paper: ChemRxiv (10.26434/chemrxiv.15007823/v1)
Code: github.com/romboai/rose-1h-nmr

Load

pip install -e ".[hub]"   # from the GitHub repo; not on PyPI

from rose import load, encode, predict

model = load()  # this repo: romboai/rose-1h-nmr
z = encode(model, spectrum, field_mhz=400.0, solvent_id=3)
clean = predict(model, noisy, task="denoise")

Input: float32 spectrum, shape (4096,) or (B, 4096), linear 0–14 ppm grid. field_mhz and solvent_id optional.

Files

File	Role
best_model.pt	pretrained checkpoint (ROSE-Pretrain-L)
rose.yaml	encoder / head config
config.json	Hub metadata

Intended use

Research and prototyping on ¹H 1D spectra: embeddings, denoising, peak maps, spectrum↔structure retrieval, coarse shift prediction. Adaptation: frozen encoder (P1) or short unfreeze (P2) — see the GitHub README.

Limitations

• Not a structure-elucidation solver. Retrieval/forward are auxiliary heads, not a replacement for assignment workflows.
• Low-field slice ($B_0$ ≤ 100 MHz) is weaker on structure-linked heads than high-field; denoise transfers better.
• Pretraining mix is multi-source (experimental + simulated). Downstream numbers need the paper holdout protocol, not a random split.

Citation

@article{diiorio2026rose,
  title   = {{ROSE}: a Foundation Model for Reusable One-dimensional
             Spectrum Embeddings in $^1$H~{NMR}},
  author  = {Di Iorio, Mattia and Mattia, Carmine and Zanda, Andrea
             and Atzori, Maurizio},
  year    = {2026},
  journal = {ChemRxiv},
  doi     = {10.26434/chemrxiv.15007823/v1},
  url     = {https://doi.org/10.26434/chemrxiv.15007823/v1},
  note    = {Preprint}
}