ROSE-1H NMR

Pretrained ¹H NMR foundation model — inference and paper adaptation protocols.

Paper: ChemRxiv. Archive: Zenodo (all versions; v0.1.0 is 10.5281/zenodo.22142632). Weights: romboai/rose-1h-nmr. Code: this repo.

Frozen [CLS] embeddings (UMAP). From the paper.

Quick start

git clone https://github.com/romboai/rose-1h-nmr.git
cd rose-1h-nmr && pip install -e ".[hub]"

from rose import load, encode
model = load()
z = encode(model, spectrum)  # float32, shape (4096,), linear 0–14 ppm

Input may also be (B, 4096). field_mhz and solvent_id are optional (solvent_id → configs/solvent_vocab.json). Local weights: load("path/to/best_model.pt").

import numpy as np
from rose import load, encode, predict

model = load()
spectrum = np.load("spectrum.npy")     # (4096,)
z = encode(model, spectrum, field_mhz=400.0, solvent_id=3)  # cdcl3 → (B, 256)

Task heads

import numpy as np
from rose import load, predict

model = load()
spectrum = np.load("spectrum.npy")
noisy = spectrum + 0.01 * np.random.randn(*spectrum.shape).astype(np.float32)

# denoise — clean spectrum from noisy input (zero-shot)
clean = predict(model, noisy, task="denoise")
# (B, 4096)

# peak — peak probability per grid point
probs = predict(model, spectrum, task="peak")
# (B, 4096) in [0, 1]

# retrieval — contrastive spectrum ↔ SMILES (one SMILES per spectrum)
out = predict(model, spectrum, task="retrieval", smiles="CCO")
# logits_per_spec (B, B), logits_per_struct (B, B), loss

# forward — ¹H shifts from structure (SMILES; spectrum sets batch size)
out = predict(model, spectrum, task="forward", smiles="CCO")
# shifts_pred (B, 32), count_pred (B,)

# pair — similarity of aligned spectrum pairs (spectrum[i] ↔ spectrum_b[i])
out = predict(model, spectrum, task="pair", spectrum_b=spectrum)
# similarity (B,), logits (B, B)

Fine-tuning

Zero-shot uses pretrained weights as-is. For adaptation, load with eval_mode=False, then adapt + param_groups:

• freeze_encoder (alias P1) — train task head only; easy domains
• unfreeze_encoder (alias P2) — train encoders + head; hard domains

P1 — peak (frozen encoder)

import numpy as np
import torch
import torch.nn.functional as F
from rose import load
from rose.grid import ppm_axis_tensor

model = load("best_model.pt", eval_mode=False)
model.adapt(mode="freeze_encoder", task="peak")
model.print_trainable_parameters()
opt = torch.optim.Adam(model.param_groups())  # lr_head=1e-3

device = next(model.parameters()).device
spectrum = torch.from_numpy(np.load("spectrum.npy")).unsqueeze(0).to(device)
ppm = ppm_axis_tensor(spectrum.size(0), device=device, cfg=model.cfg)
peak_labels = torch.zeros_like(spectrum)  # (1, 4096) in [0, 1]

logits = model.run_task("peak", spectrum=spectrum, ppm_axis=ppm)
loss = F.binary_cross_entropy_with_logits(logits, peak_labels)
loss.backward()
opt.step()

P2 — retrieval (unfreeze encoder)

import numpy as np
import torch
from rose import load, smiles_batch
from rose.grid import ppm_axis_tensor

model = load("best_model.pt", eval_mode=False)
model.adapt(mode="unfreeze_encoder", task="retrieval")
model.print_trainable_parameters()
opt = torch.optim.Adam(model.param_groups())  # lr_head=1e-3, lr_encoder=1e-5

# contrastive loss needs batch size ≥ 2
device = next(model.parameters()).device
spectrum = torch.from_numpy(
    np.stack([np.load("spectrum_a.npy"), np.load("spectrum_b.npy")])
).to(device)
ppm = ppm_axis_tensor(spectrum.size(0), device=device, cfg=model.cfg)
mol = smiles_batch(["CCO", "c1ccccc1"]).to(device)

out = model.run_task("retrieval", spectrum=spectrum, ppm_axis=ppm, mol_batch=mol)
loss = out["loss"]
loss.backward()
opt.step()

Results

Headline numbers from the ROSE paper (ROSE-Pretrain-L, 7.8M parameters, 3.2M pretrain spectra, InChIKey-14 holdout).

Internal — native heads as pretrained, identity-disjoint held-out test (same corpus, no target-domain adaptation):

Low-field slice (B₀ ≤ 100 MHz) is weaker on structure-linked heads (retrieval 45.2%, peak F1 0.20); denoise cosine stays high (0.99).

External — comparisons use the same protocol as ROSE, not a published stack under different retrieval machinery:

NMR-Solver literature (52.9% / 67.3% Top-1 / Top-10) uses a ≈10⁸ gallery plus FAISS HNSW and set-similarity rerank — not like-for-like; Top-10 at our 30k gallery already matches theirs.

Weights: Hugging Face romboai/rose-1h-nmr.

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}
}

Data

Recipes only — no pretraining parquet. Reconstruct ROSE-Pretrain-L from the cited sources (ATTRIBUTION.md) with the paper holdout and split policy. This repo ships:

• holdout H as InChIKey-14 keys (indices/holdout/)
• split policy + catalog IDs (indices/pretrain/pretrain_l_splits.meta.json)
• literature eval ID lists (indices/benchmarks/)

Spectra stay with the original distributors. Weights are on Hugging Face, not a data dump.

Repository layout

Indices hold IDs only (no spectra). Each .meta.json documents format and split policy.

License

Code and pretrained weights: Apache-2.0. See NOTICE.

Training and evaluation spectra are not in this repository. Source names, licenses, and citations: ATTRIBUTION.md.