FairNMR500

Overview

FairNMR500 is a 500-example benchmark for molecular structure elucidation from one-dimensional experimental NMR peak tables. Each example provides a molecular formula, a solvent, a ¹H NMR peak list, a ¹³C NMR peak list, and a reference molecular structure represented as SMILES.

The benchmark corresponds to the open 500-case evaluation set described in Section 5.3 of the SpecXMaster Technical Report. The report describes this set as sampled from its NMRexp test split. For the versioned evaluation protocol, see Sections 5.1 and 5.3 of report v5.

This release contains structured peak tables. Raw FID files, continuous spectral intensities, and spectrum images are not included.

Dataset structure

The dataset has one split: test, with 500 examples. Each row corresponds to one molecule and its paired ¹H/¹³C NMR information.

Field	Description
id	Stable release identifier, such as fairnmr500_000.
fair_idx	Index retained from the existing 500-example benchmark; not an upstream NMRexp record identifier.
molecular_formula	Molecular formula supplied as input.
solvent	Reported NMR solvent.
h_nmr_peaks	List of ¹H NMR peaks, with the attributes below.
c_nmr_peaks	List of ¹³C NMR peaks; each contains delta_ppm only.
gt_smiles	Existing standardized reference SMILES used for evaluation.
gt_smiles_raw	Reference SMILES before the existing standardization, retained for traceability.

Each ¹H peak contains:

Attribute	Meaning
delta_ppm	Peak chemical shift or reported centroid, in ppm.
multiplicity	Reported splitting label, such as s, d, t, or m.
j_hz	Reported coupling constants, in Hz. An empty list means not provided, not zero.
n_hydrogens	Reported hydrogen count corresponding to the integral.
range_min_ppm, range_max_ppm	Retained peak interval; equal for a single reported position.

The ¹³C records contain chemical shifts only: intensities, integrals, multiplicities, and coupling constants are not available in this benchmark file.

Usage

Load the test split:

python -m pip install datasets

from datasets import load_dataset

test = load_dataset("dptech/FairNMR500", split="test")
example = test[0]

model_input = {
    key: example[key]
    for key in ("molecular_formula", "solvent", "h_nmr_peaks", "c_nmr_peaks")
}
reference = example["gt_smiles"]

Reference structures are evaluation labels and should be withheld from the model during inference. JSONL and Parquet downloads represent the same 500 examples; the default Hugging Face configuration selects only Parquet.

Evaluation

Use rank-matcher hit@1 and hit@10, following Section 5.1 of the technical report. An example is a hit when at least one of its first k ranked predictions has the same normalized molecular key as the reference. The denominator is all 500 examples; missing, empty, or invalid predictions count as misses.

The comparison key removes atom maps and explicit hydrogens, ignores a single tetrahedral stereocenter, equates globally inverted enantiomers while preserving relative stereochemistry, and applies the documented E/Z normalization rules. Connectivity must otherwise match exactly. Predictions retain their original rank positions, including duplicate and invalid candidates.

The standalone evaluation script implements these rules. See evaluation instructions for the prediction format and command. The released GT itself is not modified using these equivalence rules.

Data source and release processing

The source dataset is NMRexp, described in NMRexp: A database of 3.3 million experimental NMR spectra. FairNMR500 is the 500-case evaluation subset used in the SpecXMaster report, not a release of the full NMRexp corpus or its full test split.

Release processing preserves sample order, reference structures, and existing spectral values. It renames fields, converts coupling-constant strings to numeric lists, removes redundant fields, and excludes stored model predictions and reranking outputs. It does not generate additional spectral measurements or correct reference structures.

The current subset does not retain per-example source DOIs, acquisition frequencies, or original-document page references.

Release validation found 500 parseable standardized reference SMILES, 500 unique canonical isomeric reference structures, and no molecular-formula discrepancies. These checks establish format and consistency; they do not constitute independent manual verification of every reference.

Citation

If you use FairNMR500, please cite the SpecXMaster Technical Report and acknowledge NMRexp as the underlying data source.

@misc{ge2026specxmaster,
  title = {SpecXMaster Technical Report},
  author = {Yutang Ge and Yaning Cui and Hanzheng Li and Jun-Jie Wang and Fanjie Xu and Jinhan Dong and Yongqi Jin and Dongxu Cui and Peng Jin and Guojiang Zhao and Hengxing Cai and Tianci Yangfeng and Xueqing Chen and Hongshuai Wang and Rong Zhu and Linfeng Zhang and Xiaohong Ji and Zhifeng Gao},
  year = {2026},
  eprint = {2603.23101},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG},
  url = {https://arxiv.org/abs/2603.23101}
}

License

This release is distributed under Creative Commons Attribution 4.0 International (CC BY 4.0), following the license declared for the underlying NMRexp dataset. See LICENSE.

Attribution: NMRexp, Junjie Wang, DOI: 10.5281/zenodo.16809534. This benchmark subset and its release formatting are provided by DP Technology. Changes from the existing benchmark records are described in the processing section above.