RxnMiner
Collection
Chemical Reaction Parsing • 3 items • Updated • 8
Datasets:
ChemRxn-Bench
🚀 Benchmark | 💻 GitHub | 📄 Paper
A benchmark for extracting complete, structured chemical reactions from literature figures: reaction schemes, substrate-scope tables with R-group templates, condition-optimization tables, multistep routes and total syntheses. Each figure comes with a manually annotated and expert-reviewed reference answer listing every concrete reaction drawn or implied in the figure, with exact molecular identities (SMILES, stereochemistry included), conditions and yields.
It accompanies the paper RxnMiner: A Multimodal Agent for Structured Reaction Data Extraction from Chemistry Figures (arXiv:2611.02826). Code: github.com/rxnminer/RxnMiner. The silver reaction corpus extracted by RxnMiner is released as UniParser/RxnMiner-Dataset.
| Figures | 603 |
| Concrete reactions | 8,510 (8,547 annotated records before the scorer's normalisation) |
| Generic templates (not scored) | 675 |
| Layouts | substrate scope 425, table 161, single step 6, complex figure 6, multistep / total synthesis 5 |
| Difficulty | Very Hard 435, Hard 37, Medium 25, Easy 28, Very Easy 78 |
Data fields
| Field | Type | Description |
|---|---|---|
id |
string | Figure id (stable across releases) |
image |
image | The figure, including its caption, table text and footnotes; the only model input |
annotation |
string (JSON) | Reference answer in the unified JSON format (below) |
layout_type |
string | substrate_scope, table, single_step, complex_figure, multistep_or_total_synthesis |
difficulty |
string | Difficulty tier, defined by how well Qwen3.8-Flash-Next performs on the figure (see the paper) |
num_reactions |
int | Concrete reactions in annotation.selected_examples |
num_templates |
int | Generic templates in annotation.reaction_templates (never scored) |
source |
string | Journal the figure was published in, or the public dataset it was collected from (ChemEAGLE, OpenChemIE) |
Unified JSON (reference answer and expected prediction format)
{
"reaction_templates": [ ... generic schemes with R groups (not scored) ... ],
"selected_examples": [
{
"instance_id": "3a",
"reactants": [{"identifier": "1a", "smiles": "...", "bbox": [x0, y0, x1, y1]}],
"products": [{"identifier": "3a", "smiles": "..."}],
"conditions": [[{"role": "catalyst", "text": "Pd(PPh3)4 (5 mol%)"}, {"role": "solvent", "text": "THF"}]],
"metrics": [{"role": "yield", "text": "82%"}]
}
]
}
In the reference answers, molecules store their structure as E-SMILES (MolParser extended SMILES) or SMILES;
predictions may use smiles, E-SMILES or SMILES (the scorer reads them in that order). Reagents printed over an
arrow are recorded as condition text. Multistep routes list their stages under steps. Other fields in the reference
answers (descriptions, footnotes, notes on conditions) document where each value comes from in the figure and are not
scored.
Evaluation
The scorer is scorer/score_reaction_strict.py in this repository. It requires rdkit==2022.9.5; other RDKit
versions can change canonical SMILES and therefore scores.
import importlib.util, json
from datasets import load_dataset
from huggingface_hub import hf_hub_download
path = hf_hub_download("UniParser/RxnMiner-Benchmark", "scorer/score_reaction_strict.py", repo_type="dataset")
spec = importlib.util.spec_from_file_location("score_reaction_strict", path)
S = importlib.util.module_from_spec(spec)
spec.loader.exec_module(S)
ds = load_dataset("UniParser/RxnMiner-Benchmark", split="test")
row = ds[0]
prediction = {...} # your system's unified JSON for row["image"]
result = S.score_both(prediction, json.loads(row["annotation"]))
print(result["reaction_f1_iso"]) # stereochemistry-aware; reaction_f1_ns ignores stereochemistry
From the root of a local copy of this repository: python scorer/score_reaction_strict.py prediction.json reference.json.
The concrete reactions in selected_examples are scored; each step of a multistep route counts as a separate
reaction, and generic templates in reaction_templates are not scored. Reaction recall = fraction of reference
products recovered; reaction precision = predicted reactions whose reactant set and product set both match a
reference reaction exactly (after RDKit canonicalisation), divided by all predicted reactions; reaction F1 is
computed per figure and averaged over figures (macro). A failed or missing prediction scores 0 and stays in the
denominator. In the paper, RxnMiner with Qwen3.8-Flash-Next reaches 71.0 macro reaction F1 (stereochemistry-aware),
versus 21.0 for the same model prompted directly.
License and use
The dataset is released under CC BY-NC-SA 4.0: non-commercial use, with attribution, and derivatives under the same license. The figures are reproduced from published articles for non-commercial research and remain the property of their publishers. Access requests are reviewed by the maintainers.
Citation
@misc{hu2026rxnminer,
title = {RxnMiner: A Multimodal Agent for Structured Reaction Data Extraction from Chemistry Figures},
author = {Hu, Jinbo and Zhang, Songren and Fu, Chuanxun and Huang, Jinrui and Wong, Zhen Hao and Cui, Dongxu and
Fan, Haoran and Ge, Muhao and Jian, Zhen and Zhou, Jian and Luo, Chengxiang and Lyu, Han and Tao, Haoyi and
Huang, Chaozheng and Li, Hanzheng and Wang, ZiChen and Ma, Hengtian and Dai, Jinyang and Wang, Nan and
Hong, Yanhui and Fang, Xi and Ke, Guolin},
year = {2026},
eprint = {2611.02826},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.02826}
}
- Downloads last month
- 12