MolP2SBench

Can a language model use molecular properties to recover a missing structural fragment? MolP2SBench presents a masked molecular scaffold and two to four valid candidate structures. Controlled changes to the supplied properties test whether predictions respond to the evidence.

Paper · Code and original release · RSI4Science usage guide

Load a task view

The task and label configurations can be loaded separately:

from datasets import load_dataset

repo = "Joshua1998/MolP2SBench"
tasks = load_dataset(repo, "dipole", split="correct")
labels = load_dataset(repo, "dipole_labels", split="correct")

# Give the model task fields only. Join labels by record_id when scoring.
answers = {row["record_id"]: row["label"] for row in labels}
print(tasks[0])

Task files and answer files remain separate. The answer is a zero-based index into candidate_completed_smiles; there is no reference reasoning trace. Labels are public, so this release is not a sealed evaluation set.

Contents

Property configurations	What is supplied
dipole, gap	Dipole magnitude; HOMO–LUMO gap
infrared, raman, ultraviolet_visible	Vibrational and electronic spectral evidence
proton_nmr_peaks, proton_nmr_integration, proton_nmr_atom	Three representations of proton magnetic-shielding information
carbon_nmr_peaks, carbon_nmr_degeneracy, carbon_nmr_atom	Three representations of carbon magnetic-shielding information
electronic, magnetic_resonance, vibrational	Property-family bundles
all_properties	Seven property types together

Each configuration has three condition views, exposed as Hugging Face split names:

View	Evidence	Evaluation purpose
correct	Properties associated with the source structure	Recovery with matched evidence
shuffled	Properties from another molecule, with the original label retained	Sensitivity to disrupted property–structure pairing
counterfactual	Properties of another valid completion on the same scaffold	Recovery directed toward the donor completion

These are paired experimental conditions, not independent train/validation/test partitions. Each channel has 136 correct and 136 shuffled records. Counterfactual views contain 240 or 258 records, depending on the channel. There are 7,824 task rows across all views and channels; repeated views are not additional independent molecules. Every task configuration has a corresponding <configuration>_labels configuration.

Fields and scoring

Tasks contain record_id, masked_smiles, mask, properties, and candidate_completed_smiles. Property metadata records units and representations; NMR fields describe calculated magnetic shieldings, not a general set of experimental chemical shifts. Answer files contain record_id and label.

The original prompts and Python scorer are included. Run from a downloaded repository snapshot:

python scripts/score_tasks.py --property dipole --condition correct --predictions predictions.jsonl

The scorer reports exact-match outcomes and invalid or missing predictions. It does not score free-text reasoning, infer why a model failed, or implement the paper's family-weighted PGR/CFR metrics. See SOURCE_README.md for the original prediction schema and protocol.

Using the data for improvement research

Compare outcomes by property and evidence condition to form testable hypotheses about evidence use. A change in one answer is not a diagnosis of an internal reasoning mechanism. Evaluate revisions on independent molecular families under matched budgets; the views here do not supply an independent held-out split. This release supports diagnostics and experiments, and does not establish that recursive self-improvement or open-ended molecular discovery has been achieved.

Source and permissions

The data, prompts and scripts are copied without changes from GitHub commit 1ede82bf68f634097b296b1937d66a9b04bbf269. source-manifest.json records their SHA-256 hashes. Dataset construction and preprocessing are described in the linked paper. The upstream sources retain their respective terms:

• QM9: HOMO–LUMO gaps; data record, CC0.
• QM9S: dipole, IR, Raman and UV–Vis properties; data record, CC BY 4.0; source article.
• QM9NMR: gas-phase computed magnetic shieldings; data record, CC BY 4.0; project and citation information.

This mirror preserves the released task payloads and does not redistribute the complete upstream databases. The benchmark constructs masked-scaffold candidate-selection tasks and condition-specific property views from those sources; the upstream measurements and calculations are not newly generated here.

The MolP2SBench benchmark data, task formulations, prompts and accompanying documentation in this release are licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). Reuse and adaptation, including commercial use, are permitted with attribution, a link to the license and an indication of changes. See LICENSE.

The upstream data retain the terms listed above; this release does not impose new restrictions on CC0 material. Please credit MolP2SBench and the relevant upstream sources when reusing their material. The original Python scripts in scripts/ are included unchanged for reference; this data license does not assign them a software license.

Citation

@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}
}