Huggggooo commited on
Commit
e4bcd5b
·
verified ·
1 Parent(s): 58e89f5

Release FairNMR500 v1.0.0

Browse files

Publish 500 paired 1H/13C NMR peak-table examples with molecular formula, solvent, reference SMILES, documentation, and standalone rank-matcher evaluation.

Files changed (7) hide show
  1. LICENSE +19 -0
  2. README.md +151 -0
  3. data/test.jsonl +0 -0
  4. data/test.parquet +3 -0
  5. evaluation/README.md +48 -0
  6. evaluation/metric.py +209 -0
  7. requirements.txt +2 -0
LICENSE ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ FairNMR500
2
+
3
+ This release is licensed under the Creative Commons Attribution 4.0
4
+ International license (CC BY 4.0).
5
+
6
+ License summary: https://creativecommons.org/licenses/by/4.0/
7
+ Full legal text: https://creativecommons.org/licenses/by/4.0/legalcode
8
+
9
+ Underlying data attribution:
10
+ NMRexp: A database of 3.37 million experimental NMR spectra
11
+ Creator: Junjie Wang
12
+ Dataset DOI: https://doi.org/10.5281/zenodo.16809534
13
+ Source license declaration: https://zenodo.org/api/records/16809534
14
+
15
+ The FairNMR500 benchmark subset, release formatting, and accompanying evaluation
16
+ code are provided by DP Technology. Release formatting renames fields, parses
17
+ coupling-constant lists, and removes redundant fields and stored model outputs.
18
+ The release does not modify the existing reference structures or spectral
19
+ measurements. Please retain attribution and indicate any further modifications.
README.md CHANGED
@@ -1,3 +1,154 @@
1
  ---
 
2
  license: cc-by-4.0
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ pretty_name: FairNMR500
3
  license: cc-by-4.0
4
+ size_categories:
5
+ - n<1K
6
+ tags:
7
+ - chemistry
8
+ - nmr
9
+ - molecular-structure-elucidation
10
+ - benchmark
11
+ - specxmaster
12
+ configs:
13
+ - config_name: default
14
+ data_files:
15
+ - split: test
16
+ path: data/test.parquet
17
  ---
18
+
19
+ # FairNMR500
20
+
21
+ ## Overview
22
+
23
+ FairNMR500 is a 500-example benchmark for molecular structure elucidation from
24
+ one-dimensional experimental NMR peak tables. Each example provides a molecular
25
+ formula, a solvent, a ¹H NMR peak list, a ¹³C NMR peak list, and a reference molecular
26
+ structure represented as SMILES.
27
+
28
+ The benchmark corresponds to the open 500-case evaluation set described in
29
+ Section 5.3 of the [SpecXMaster Technical Report](https://arxiv.org/abs/2603.23101).
30
+ The report describes this set as sampled from its NMRexp test split. For the
31
+ versioned evaluation protocol, see
32
+ [Sections 5.1 and 5.3 of report v5](https://arxiv.org/html/2603.23101v5#S5.SS1).
33
+
34
+ This release contains structured peak tables. Raw FID files, continuous spectral
35
+ intensities, and spectrum images are not included.
36
+
37
+ ## Dataset structure
38
+
39
+ The dataset has one split: **test**, with **500 examples**. Each row corresponds
40
+ to one molecule and its paired ¹H/¹³C NMR information.
41
+
42
+ | Field | Description |
43
+ | --- | --- |
44
+ | `id` | Stable release identifier, such as `fairnmr500_000`. |
45
+ | `fair_idx` | Index retained from the existing 500-example benchmark; not an upstream NMRexp record identifier. |
46
+ | `molecular_formula` | Molecular formula supplied as input. |
47
+ | `solvent` | Reported NMR solvent. |
48
+ | `h_nmr_peaks` | List of ¹H NMR peaks, with the attributes below. |
49
+ | `c_nmr_peaks` | List of ¹³C NMR peaks; each contains `delta_ppm` only. |
50
+ | `gt_smiles` | Existing standardized reference SMILES used for evaluation. |
51
+ | `gt_smiles_raw` | Reference SMILES before the existing standardization, retained for traceability. |
52
+
53
+ Each ¹H peak contains:
54
+
55
+ | Attribute | Meaning |
56
+ | --- | --- |
57
+ | `delta_ppm` | Peak chemical shift or reported centroid, in ppm. |
58
+ | `multiplicity` | Reported splitting label, such as `s`, `d`, `t`, or `m`. |
59
+ | `j_hz` | Reported coupling constants, in Hz. An empty list means not provided, not zero. |
60
+ | `n_hydrogens` | Reported hydrogen count corresponding to the integral. |
61
+ | `range_min_ppm`, `range_max_ppm` | Retained peak interval; equal for a single reported position. |
62
+
63
+ The ¹³C records contain chemical shifts only: intensities, integrals,
64
+ multiplicities, and coupling constants are not available in this benchmark file.
65
+
66
+ ## Usage
67
+
68
+ Load the test split:
69
+
70
+ ```bash
71
+ python -m pip install datasets
72
+ ```
73
+
74
+ ```python
75
+ from datasets import load_dataset
76
+
77
+ test = load_dataset("dptech/FairNMR500", split="test")
78
+ example = test[0]
79
+
80
+ model_input = {
81
+ key: example[key]
82
+ for key in ("molecular_formula", "solvent", "h_nmr_peaks", "c_nmr_peaks")
83
+ }
84
+ reference = example["gt_smiles"]
85
+ ```
86
+
87
+ Reference structures are evaluation labels and should be withheld from the
88
+ model during inference. JSONL and Parquet downloads represent the same 500
89
+ examples; the default Hugging Face configuration selects only Parquet.
90
+
91
+ ## Evaluation
92
+
93
+ Use **rank-matcher hit@1 and hit@10**, following Section 5.1 of the technical
94
+ report. An example is a hit when at least one of its first k ranked predictions
95
+ has the same normalized molecular key as the reference. The denominator is all
96
+ 500 examples; missing, empty, or invalid predictions count as misses.
97
+
98
+ The comparison key removes atom maps and explicit hydrogens, ignores a single
99
+ tetrahedral stereocenter, equates globally inverted enantiomers while preserving
100
+ relative stereochemistry, and applies the documented E/Z normalization rules.
101
+ Connectivity must otherwise match exactly. Predictions retain their original
102
+ rank positions, including duplicate and invalid candidates.
103
+
104
+ The standalone [evaluation script](evaluation/metric.py) implements these rules.
105
+ See [evaluation instructions](evaluation/README.md) for the prediction format and
106
+ command. The released GT itself is not modified using these equivalence rules.
107
+
108
+ ## Data source and release processing
109
+
110
+ The source dataset is **NMRexp**, described in
111
+ [NMRexp: A database of 3.3 million experimental NMR spectra](https://doi.org/10.1038/s41597-025-06245-5).
112
+ FairNMR500 is the 500-case evaluation subset used in the SpecXMaster report, not
113
+ a release of the full NMRexp corpus or its full test split.
114
+
115
+ Release processing preserves sample order, reference structures, and existing
116
+ spectral values. It renames fields, converts coupling-constant strings to numeric
117
+ lists, removes redundant fields, and excludes stored model predictions and
118
+ reranking outputs. It does not generate additional spectral measurements or
119
+ correct reference structures.
120
+
121
+ The current subset does not retain per-example source DOIs, acquisition
122
+ frequencies, or original-document page references.
123
+
124
+ Release validation found 500 parseable standardized
125
+ reference SMILES, 500 unique canonical isomeric reference structures, and no
126
+ molecular-formula discrepancies. These checks establish format and consistency;
127
+ they do not constitute independent manual verification of every reference.
128
+
129
+ ## Citation
130
+
131
+ If you use FairNMR500, please cite the SpecXMaster Technical Report and acknowledge
132
+ NMRexp as the underlying data source.
133
+
134
+ ```bibtex
135
+ @misc{ge2026specxmaster,
136
+ title = {SpecXMaster Technical Report},
137
+ 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},
138
+ year = {2026},
139
+ eprint = {2603.23101},
140
+ archivePrefix = {arXiv},
141
+ primaryClass = {cs.LG},
142
+ url = {https://arxiv.org/abs/2603.23101}
143
+ }
144
+ ```
145
+
146
+ ## License
147
+
148
+ This release is distributed under **Creative Commons Attribution 4.0
149
+ International (CC BY 4.0)**, following the license declared for the underlying
150
+ [NMRexp dataset](https://zenodo.org/records/16809534). See [LICENSE](LICENSE).
151
+
152
+ Attribution: NMRexp, Junjie Wang, DOI: 10.5281/zenodo.16809534. This benchmark
153
+ subset and its release formatting are provided by DP Technology. Changes from
154
+ the existing benchmark records are described in the processing section above.
data/test.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
data/test.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:71dde01aab49cee39f77252e94e31580fbe0e717cc4a13db4b7c98ee4470ad6d
3
+ size 83120
evaluation/README.md ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # FairNMR500 evaluation
2
+
3
+ Requires Python 3.10+ and RDKit. From the repository root:
4
+
5
+ ```bash
6
+ python -m pip install -r requirements.txt
7
+ python evaluation/metric.py --predictions predictions.jsonl --output metrics.json
8
+ ```
9
+
10
+ Prepare one JSON object per line, using the released sample `id` and an ordered
11
+ list of predicted SMILES:
12
+
13
+ ```json
14
+ {"id": "fairnmr500_000", "smiles": ["CCO", "CCCO"]}
15
+ {"id": "fairnmr500_001", "smiles": []}
16
+ ```
17
+
18
+ These two rows illustrate the format, not benchmark predictions. Empty lists,
19
+ omitted samples, empty strings, and invalid SMILES count as misses. Invalid and
20
+ duplicate SMILES retain their original candidate positions; candidates are not
21
+ deduplicated or promoted. Duplicate IDs, unknown IDs, malformed JSON, and
22
+ non-string candidates are rejected as input errors. Prediction order across
23
+ rows does not matter; candidate order within each row does.
24
+
25
+ The script reports `correct@1`, `correct@10`, `hit@1`, and `hit@10`. The hit rates
26
+ are fractions between 0 and 1. The denominator is the full reference file
27
+ (500 examples for this release), including samples without predictions.
28
+
29
+ ## Molecular comparison
30
+
31
+ The rank matcher follows Section 5.1 of the
32
+ [SpecXMaster Technical Report, v5](https://arxiv.org/html/2603.23101v5#S5.SS1):
33
+
34
+ - Parse SMILES with RDKit, remove atom maps and explicit hydrogens, and produce
35
+ canonical isomeric representations.
36
+ - Ignore the configuration of a single tetrahedral stereocenter.
37
+ - For multiple tetrahedral stereocenters, equate full-inversion enantiomers and
38
+ preserve relative stereochemistry; partial inversion is not generally equal.
39
+ - Normalize unspecified stereogenic double bonds to E, while retaining explicit
40
+ E/Z labels except for the original matcher's nitrogen-associated patterns.
41
+ The exact implemented pattern is a double-bond endpoint that is nitrogen, or
42
+ nitrogen directly attached to either endpoint.
43
+ - Require all other connectivity to match exactly. Empty/invalid SMILES have no
44
+ comparison key and cannot match a valid reference.
45
+
46
+ The evaluation equivalences do not change the GT stored in the dataset. The
47
+ script uses no internal services, model weights, or inference-project modules.
48
+ Dependencies are pinned to the RDKit version used in release validation.
evaluation/metric.py ADDED
@@ -0,0 +1,209 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """Standalone rank-matcher hit@k for FairNMR500 (Python >= 3.10).
3
+
4
+ The structural normalization follows the SpecXMaster rank matcher. Candidate
5
+ positions are preserved, including duplicates and invalid SMILES.
6
+ """
7
+
8
+ from __future__ import annotations
9
+
10
+ import argparse
11
+ import json
12
+ from functools import lru_cache
13
+ from pathlib import Path
14
+
15
+ from rdkit import Chem, rdBase
16
+
17
+
18
+ def _canonical_no_map_no_h(mol):
19
+ mol = Chem.Mol(mol)
20
+ for atom in mol.GetAtoms():
21
+ atom.SetAtomMapNum(0)
22
+ mol = Chem.RemoveHs(mol, implicitOnly=False)
23
+ return Chem.MolToSmiles(mol, canonical=True, isomericSmiles=True)
24
+
25
+
26
+ def _find_chiral_center_indices(mol):
27
+ mol = Chem.Mol(mol)
28
+ Chem.AssignStereochemistry(mol, cleanIt=True, force=True)
29
+ centers = Chem.FindMolChiralCenters(
30
+ mol, includeUnassigned=True, useLegacyImplementation=False
31
+ )
32
+ return sorted({int(idx) for idx, _ in centers})
33
+
34
+
35
+ def _clear_single_chiral_center(mol, center_idx):
36
+ mol = Chem.Mol(mol)
37
+ atom = mol.GetAtomWithIdx(int(center_idx))
38
+ atom.SetChiralTag(Chem.rdchem.ChiralType.CHI_UNSPECIFIED)
39
+ if atom.HasProp("_CIPCode"):
40
+ atom.ClearProp("_CIPCode")
41
+ return mol
42
+
43
+
44
+ def _invert_all_tetrahedral_centers(mol, center_indices):
45
+ mol = Chem.Mol(mol)
46
+ for idx in center_indices:
47
+ atom = mol.GetAtomWithIdx(int(idx))
48
+ tag = atom.GetChiralTag()
49
+ if tag == Chem.rdchem.ChiralType.CHI_TETRAHEDRAL_CW:
50
+ atom.SetChiralTag(Chem.rdchem.ChiralType.CHI_TETRAHEDRAL_CCW)
51
+ elif tag == Chem.rdchem.ChiralType.CHI_TETRAHEDRAL_CCW:
52
+ atom.SetChiralTag(Chem.rdchem.ChiralType.CHI_TETRAHEDRAL_CW)
53
+ return mol
54
+
55
+
56
+ def _alkene_has_terminal_nitrogen(bond):
57
+ # Preserve the original matcher's exact pattern: a double-bond endpoint
58
+ # is N, or N is directly attached to either endpoint.
59
+ begin, end = bond.GetBeginAtom(), bond.GetEndAtom()
60
+ if begin.GetAtomicNum() == 7 or end.GetAtomicNum() == 7:
61
+ return True
62
+ return any(
63
+ neighbor.GetIdx() != other.GetIdx() and neighbor.GetAtomicNum() == 7
64
+ for atom, other in ((begin, end), (end, begin))
65
+ for neighbor in atom.GetNeighbors()
66
+ )
67
+
68
+
69
+ def _clear_neighbor_single_bond_dirs(double_bond):
70
+ for atom in (double_bond.GetBeginAtom(), double_bond.GetEndAtom()):
71
+ for bond in atom.GetBonds():
72
+ if (
73
+ bond.GetIdx() != double_bond.GetIdx()
74
+ and bond.GetBondType() == Chem.BondType.SINGLE
75
+ ):
76
+ bond.SetBondDir(Chem.BondDir.NONE)
77
+
78
+
79
+ def _force_unspecified_alkene_to_trans(mol):
80
+ mol = Chem.Mol(mol)
81
+ Chem.FindPotentialStereoBonds(mol)
82
+ changed = False
83
+ for bond in mol.GetBonds():
84
+ if bond.GetBondType() != Chem.BondType.DOUBLE or bond.GetIsAromatic():
85
+ continue
86
+ if _alkene_has_terminal_nitrogen(bond):
87
+ if bond.GetStereo() != Chem.BondStereo.STEREONONE:
88
+ bond.SetStereo(Chem.BondStereo.STEREONONE)
89
+ _clear_neighbor_single_bond_dirs(bond)
90
+ changed = True
91
+ continue
92
+ if bond.GetStereo() == Chem.BondStereo.STEREOANY:
93
+ bond.SetStereo(Chem.BondStereo.STEREOE)
94
+ changed = True
95
+ if changed:
96
+ Chem.SetDoubleBondNeighborDirections(mol)
97
+ return mol
98
+
99
+
100
+ @lru_cache(maxsize=65536)
101
+ def get_rank_key(smiles: str) -> str | None:
102
+ """Return a molecular comparison key; empty/invalid input has no key."""
103
+ if not isinstance(smiles, str) or not smiles.strip():
104
+ return None
105
+ with rdBase.BlockLogs():
106
+ try:
107
+ mol = Chem.MolFromSmiles(smiles.strip())
108
+ if mol is None:
109
+ return None
110
+ mol = _force_unspecified_alkene_to_trans(mol)
111
+ centers = _find_chiral_center_indices(mol)
112
+ if not centers:
113
+ return _canonical_no_map_no_h(mol)
114
+ if len(centers) == 1:
115
+ return _canonical_no_map_no_h(
116
+ _clear_single_chiral_center(mol, centers[0])
117
+ )
118
+ own = _canonical_no_map_no_h(mol)
119
+ mirror = _canonical_no_map_no_h(
120
+ _invert_all_tetrahedral_centers(mol, centers)
121
+ )
122
+ return "||".join(sorted([own, mirror]))
123
+ except Exception:
124
+ return None
125
+
126
+
127
+ def _index_unique(rows, label):
128
+ indexed = {}
129
+ for row in rows:
130
+ if not isinstance(row, dict) or not isinstance(row.get("id"), str):
131
+ raise ValueError(f"{label}: each row must contain a string id")
132
+ if not row["id"] or row["id"] in indexed:
133
+ raise ValueError(f"{label}: empty or duplicate id {row['id']!r}")
134
+ indexed[row["id"]] = row
135
+ return indexed
136
+
137
+
138
+ def evaluate(references, predictions, ks=(1, 10)):
139
+ """Score all reference IDs; omitted prediction rows count as misses."""
140
+ if not ks or any(type(k) is not int or k < 1 for k in ks):
141
+ raise ValueError("ks must contain positive integer cutoffs")
142
+ refs = _index_unique(references, "references")
143
+ preds = _index_unique(predictions, "predictions")
144
+ if not refs:
145
+ raise ValueError("reference dataset is empty")
146
+ unknown = set(preds) - set(refs)
147
+ if unknown:
148
+ raise ValueError(f"unknown prediction IDs: {sorted(unknown)[:5]}")
149
+ for sample_id, row in preds.items():
150
+ candidates = row.get("smiles")
151
+ if not isinstance(candidates, list) or not all(
152
+ isinstance(s, str) for s in candidates
153
+ ):
154
+ raise ValueError(f"{sample_id}: smiles must be a list of strings")
155
+
156
+ counts = {k: 0 for k in ks}
157
+ for sample_id, ref in refs.items():
158
+ gt = ref.get("gt_smiles")
159
+ if not isinstance(gt, str) or not (gt_key := get_rank_key(gt)):
160
+ raise ValueError(f"{sample_id}: invalid reference SMILES")
161
+ candidates = preds.get(sample_id, {}).get("smiles", [])
162
+ rank = next(
163
+ (i for i, smi in enumerate(candidates[: max(ks)], start=1)
164
+ if get_rank_key(smi) == gt_key),
165
+ None,
166
+ )
167
+ for k in ks:
168
+ counts[k] += int(rank is not None and rank <= k)
169
+
170
+ result = {
171
+ "num_samples": len(refs),
172
+ "num_prediction_rows": len(preds),
173
+ "num_missing_predictions": len(refs) - len(preds),
174
+ "match_mode": "rank-matcher",
175
+ "rdkit_version": rdBase.rdkitVersion,
176
+ }
177
+ for k in ks:
178
+ result[f"correct@{k}"] = counts[k]
179
+ result[f"hit@{k}"] = counts[k] / len(refs)
180
+ return result
181
+
182
+
183
+ def read_jsonl(path):
184
+ with Path(path).open(encoding="utf-8") as handle:
185
+ return [json.loads(line) for line in handle if line.strip()]
186
+
187
+
188
+ def main():
189
+ parser = argparse.ArgumentParser(description=__doc__)
190
+ parser.add_argument(
191
+ "--references", type=Path,
192
+ default=Path(__file__).resolve().parents[1] / "data/test.jsonl",
193
+ )
194
+ parser.add_argument("--predictions", type=Path, required=True)
195
+ parser.add_argument("--output", type=Path)
196
+ args = parser.parse_args()
197
+ result = evaluate(read_jsonl(args.references), read_jsonl(args.predictions))
198
+ text = json.dumps(result, indent=2, ensure_ascii=False) + "\n"
199
+ print(text, end="")
200
+ if args.output:
201
+ if args.output.resolve() in {
202
+ args.references.resolve(), args.predictions.resolve()
203
+ }:
204
+ raise ValueError("output cannot overwrite an input file")
205
+ args.output.write_text(text, encoding="utf-8")
206
+
207
+
208
+ if __name__ == "__main__":
209
+ main()
requirements.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ # Python >= 3.10; needed only for evaluation/metric.py.
2
+ rdkit==2025.9.3