Datasets:
Canis M token pack
12,000 public molecules, each carrying a quantum token: a fixed-length representation of the molecule's correlated electronic structure, computed by our engine and read by Canis M in place of atomic coordinates.
The comparison that matters takes an afternoon: take your own descriptor, concatenate our token, retrain your model, and measure what changed. Every number on the model card was produced that way.
What is in the file
token_pack.parquet, 12,000 rows, one per molecule.
| Column | Type | What it is |
|---|---|---|
canonical_smiles |
string | the molecule |
shard |
string | which subset it was drawn from |
n_heavy |
int16 | heavy-atom count |
token |
list[float] × 648 | the quantum token, in the published basis |
corr_residual_hartree |
double | the reference label, E_CASCI(6,6)/6-31G - E_MP2 |
mr_index |
double | multireference character, a difficulty measure |
Coverage is 10, 11 and 12 heavy atoms, 4,000 molecules each, drawn from public GDB-13 at the relaxed geometry, which is the geometry Canis M is defined on. Elements are C, N, O, S and Cl.
Using it
import numpy as np, pyarrow.parquet as pq, torch, torch.nn as nn
from huggingface_hub import hf_hub_download
t = pq.read_table(hf_hub_download("SiriusQuantum/canis-m-token-pack",
"token_pack.parquet", repo_type="dataset"))
X = np.stack(t["token"].to_numpy(zero_copy_only=False)) # (12000, 648) float32
y = t["corr_residual_hartree"].to_numpy() # Hartree
ck = torch.load(hf_hub_download("SiriusQuantum/canis-m", "canis-m-v1.pt"), map_location="cpu")
net = nn.Sequential(nn.Linear(ck["in_dim"], ck["hid"]), nn.SiLU(),
nn.Linear(ck["hid"], ck["hid"]), nn.SiLU(),
nn.Linear(ck["hid"], ck["n_out"]))
net.load_state_dict(ck["state_dict"]); net.eval()
with torch.no_grad():
pred = net(torch.from_numpy(X))[:, 0].numpy() # column 0 is the correction, in Hartree
print(f"MAE {1000 * abs(pred - y).mean():.4f} mHa on {len(y):,} molecules")
MAE 0.2022 mHa on 12,000 molecules
That is the whole loop: tokens in, correction out, scored against the reference label. Swap net
for your own model and X for your own descriptor with these tokens concatenated, and you have
the comparison this pack exists for.
The tokens are ready to use as they are: do not standardise them again, and do not rescale the labels, which are in Hartree.
What it is not
These are public, enumerated molecules, not anybody's compound library.
Your own molecules are not in here. Their tokens have to be minted. Send us the list and we will mint them.
Drug-sized molecules are a different model. On 40,600 real compounds of 20 to 52 heavy atoms, quantum tokens put 93.7 percent inside chemical accuracy where the strongest classical descriptor we tested reaches 65.0 percent. That model is not released. If your molecules are that size, let's talk.
Citation
@article{karli2026tokenising,
title = {Tokenising quantum data for label-efficient training of molecular models},
author = {Karli, Derya},
journal = {ChemRxiv},
year = {2026},
doi = {10.26434/chemrxiv.15009119/v1},
url = {https://doi.org/10.26434/chemrxiv.15009119/v1}
}
Contact
TokenService access, and minting for your own models: info@siriusquantum.com
Model, integration and bug reports: dev@siriusquantum.com
- Downloads last month
- 18