SOURCE_README.md

MolP2SBench

MolP2SBench

Usual molecular benchmarks give a complete structure and ask for a property. MolP2SBench studies the inverse problem: a fragment of the molecule is hidden, and the system must decide which chemically valid completion fits the remaining graph and the supplied physical evidence.

The work is a test of property-grounded completion. A high recovery rate can come from valence, common functional-group templates, or SMILES familiarity. The released views ask a narrower question: when several completions are already valid on the same scaffold, does the supplied property change the selection in the direction the evidence implies?

Task

The task is defined on molecular graphs and shown as SMILES. A connected functional group is replaced by a mask token. The system receives:

• the masked SMILES and mask metadata
• a short list of completed candidate SMILES, all valid on that scaffold
• the properties listed for that directory

It returns one candidate index. The closed comparison supplies that task text only: no tools, and no web search or literature retrieval.

Artifact	Path
Single-property prompt	prompts/single_property_system.txt, prompts/single_property_user.txt
Bundle prompt	prompts/property_bundle_system.txt, prompts/property_bundle_user.txt
All-properties prompt	prompts/all_properties_system.txt, prompts/all_properties_user.txt
Tasks	data/<property>/{correct,shuffled,counterfactual}.jsonl
Labels	data/<property>/{correct,shuffled,counterfactual}.labels.jsonl

Each task line is one JSON object with record_id, masked_smiles, mask, properties, and candidate_completed_smiles. The answer is not in that object. It is the sibling file <condition>.labels.jsonl. A label line has record_id and label. label is the 0-based index of the target candidate.

The returned object matches the prompt:

{
  "record_id": "<input record_id>",
  "selected_candidate_index": 0,
  "reasoning": "<concise justification>"
}

selected_candidate_index is a 0-based index into candidate_completed_smiles, in the order given.

Single-property directories carry one representation. A bundle carries one family. all_properties carries seven properties: the dipole norm, the gap, infrared, Raman, ultraviolet–visible, proton NMR with integration, and carbon NMR with degeneracy.

Directory	Property
dipole	Dipole moment, Euclidean norm
gap	Gap between the highest occupied and lowest unoccupied molecular orbitals
infrared	Infrared spectrum
raman	Raman spectrum
ultraviolet_visible	Ultraviolet–visible excitations
proton_nmr_peaks	Proton NMR, peaks only
proton_nmr_integration	Proton NMR, peaks with integration
proton_nmr_atom	Proton NMR, atom-wise
carbon_nmr_peaks	Carbon NMR, peaks only
carbon_nmr_degeneracy	Carbon NMR, peaks with degeneracy
carbon_nmr_atom	Carbon NMR, atom-wise
electronic	Dipole norm, gap, and ultraviolet–visible
magnetic_resonance	Proton NMR with integration, and carbon NMR with degeneracy
vibrational	Infrared and Raman spectra
all_properties	The seven properties listed above

Correct and shuffled files have 136 records each. Counterfactual files have 240 records for dipole, infrared, Raman, ultraviolet–visible, electronic, vibrational, and all_properties. They have 258 records for the gap, each NMR representation, and magnetic_resonance.

Views

Exact match on the correct-property view is not evidence that the property was used. The same scaffolds are released under three views:

View	File	What is supplied	What it tests
Correct property	correct.jsonl	The property of the source molecule	Recovery under the matched property
Shuffled property	shuffled.jsonl	The same property type, taken from a different molecule	Whether the value is read or ignored
Counterfactual property	counterfactual.jsonl	The property of another valid completion on the same scaffold	Whether the choice moves toward that completion

Correct and shuffled views of a property share record_id and the candidate list. The shuffled file changes the property, not the answer index. A structure-only comparison uses that same masked graph and candidate list with the property omitted. The prompt states that case: if no property is supplied, use only the structural context.

A use claim needs a shuffle drop and a donor-directed counterfactual flip, scoped to one model and one property.

Environment

Loading and scoring use Python 3.9 or newer and the standard library.

python3 scripts/build_tasks.py
python3 scripts/score_tasks.py \
  --property dipole \
  --condition correct \
  --predictions predictions.jsonl

scripts/build_tasks.py loads every task file and prints the record counts. scripts/score_tasks.py compares one prediction file with the label file for that property and view. --property is a directory name from the table above. --condition is correct, shuffled, or counterfactual.

Trajectory

Score one property and one view at a time. The prediction file is JSONL, one object per line:

Field	Contents
record_id	Task id
selected_candidate_index	0-based index into candidate_completed_smiles
reasoning	Justification. The scorer does not read this field.
invalid	true when the submission does not parse. Omit the index on that line.

A missing record, a line marked invalid, or an index outside the candidate list counts as incorrect. A duplicate record_id is rejected. The scorer prints the hit count, the number of tasks, the accuracy, and the counts of missing, invalid, and out-of-range lines.