Scientific datasets

Use these datasets to test your model, inspect its errors and study improvements in your own workflow.

Phase 1 focuses on property-guided structure recovery. Start with the MolP2SBench teaching package or the original benchmark.

Our first-stage goal
Search and filter datasets

Browse by the resource’s main scientific field.

Capabilities describe what a task examines; they are separate from its scientific field.

MolP2SBenchMolecular discoveryRecover masked molecular fragments from property evidence.Teaching preview · not yet reviewed hereView detailsHide details

Recover a masked molecular fragment using computed physical properties. Known structures provide reference answers; controlled views help investigate changes in predictions.

Original paper

Teaching starter

2 observations · 2 candidates · 1 molecular family

Includes model inputs, separate labels, a scorer and a Python example. Python 3.9+; no API key needed for the smoke test.

Independent scientific review not recorded.

Teaching adaptation: source and reuse notes.

Full benchmark

136 source problems per property channel · 2–4 candidates

Use the original repository for the complete task views, labels, prompts and scorer. Its format differs from the teaching starter.

This listing documents the source version, citation and interface. Independent scientific review of the full dataset is not recorded.

External source. Check the repository and underlying data’s reuse terms.

  • Structure recovery
  • Evidence use
  • Counterfactual response

Capabilities examined in the full benchmark. The teaching pair introduces the interface.

What do these counts mean?

A case is one input record to answer. Here, a family groups one molecular structure-recovery problem from the original benchmark with all its related observations. A property channel supplies a kind of evidence, such as dipole magnitude. An evidence condition specifies how it is paired with the case: correct, shuffled or counterfactual. Keep related records together when building independent evaluation splits.

Input
Masked structure, property values and candidate completions
Output
A candidate choice for each observation
Feedback
Correctness and reference comparisons on development cases

Source and citation

Benchmark paper. Yuhang Xie, Zhiyi Shi, Zeyi Zhang, Zhisheng Qi, Yushun Dong, Yu Wang, Xiongye Xiao. MolP2SBench: Do LLMs Really Reason from Molecular Properties to Structure? arXiv:2610.12522 (2026).

Cite this resource

Cite the benchmark paper for MolP2SBench tasks and design. For QM9S-derived records, also cite the underlying QM9S paper and identify the dataset release used. Each source has its own reuse terms. A deep learning model for predicting selected organic molecular spectra · QM9S dataset release.

Select and copy the entries below, or download the BibTeX file above.

@misc{xie2026molp2sbench,
  title = {MolP2SBench: Do LLMs Really Reason from Molecular Properties to Structure?},
  author = {Yuhang Xie and Zhiyi Shi and Zeyi Zhang and Zhisheng Qi and Yushun Dong and Yu Wang and Xiongye Xiao},
  year = {2026},
  eprint = {2610.12522},
  archivePrefix = {arXiv},
  primaryClass = {physics.chem-ph},
  url = {https://arxiv.org/abs/2610.12522}
}

@article{zou2023molecularspectra,
  title = {A deep learning model for predicting selected organic molecular spectra},
  author = {Zihan Zou and Yujin Zhang and Lijun Liang and Mingzhi Wei and Jiancai Leng and Jun Jiang and Yi Luo and Wei Hu},
  journal = {Nature Computational Science},
  year = {2023},
  volume = {3},
  pages = {957--964},
  doi = {10.1038/s43588-023-00550-y},
  url = {https://doi.org/10.1038/s43588-023-00550-y}
}
Where the data lives

The teaching package is hosted here; the full benchmark remains in its source repository. Contributed datasets enter the reviewed collection after reviewers inspect the files and independently reproduce their checks.

Use the data in your improvement loop

  1. Run your model

    Give it the task inputs and collect predictions. Connect your model

  2. Learn from feedback

    Score development cases and investigate errors. Follow a feedback example

  3. Check fresh cases

    Compare versions on independent families with the same scorer and budget. Start from the full source

Use the teaching pair to check the interface. For research comparisons, group related cases before splitting and reserve independent evaluation families. Your system supplies the revisions. Explore the proposed evaluation protocol

Using a dataset in your research? Share feedback or a result

Bring your data to the collection

Propose original or curated molecular data. Acceptance requires versioned files, reproducible checks and scientific review. A paper is optional.

Contribute a dataset