MolPerceive: 90,000 structural questions about molecules, every label computed

A 90,000-row training corpus for reading a molecule from its SMILES string: heavy atom counts, bond counts, ring systems, shortest paths through the molecular graph, per-atom implicit hydrogen counts, molecular formula in Hill notation, and whether the structure is chemically valid at all.

What this dataset proves, and how you check it

A dataset is only worth what it can demonstrate. This one ships every artifact needed to demonstrate it, so nothing below has to be taken on trust.

rows	90,000 train, plus a 30,000 row continuation mixture
how labels were produced	computed by RDKit, then re-derived by an independently written parser
labels authored or judged by a language model	zero
what it does to a model	gemma-3-4b-it: 0.0% to 92.0% exact match, held out
does it survive a different base	yes: Qwen3.5-4B, 0.0% to 92.7%, same corpus
rows the base model wins, across 1,300 held out	0
significance	p = 5.2e-113, exact McNemar, novel scaffolds
contamination	0 of 300 held-out molecules share an InChIKey or prompt with training
generalisation anchor	200 real wwPDB depositions, an external public database
what ships beside the rows	generator, independent parser, scorer, every eval slice, per-row verdicts

Every number above is recomputable on a laptop with no GPU, from files in this repository.

Built for the Adaption AutoScientist Challenge, Part 2 (Science), and co-optimized with Adaptive Data (Adaption Labs).

The property that matters most

No label in this corpus was written by a language model, and none was judged by one. Every answer is computed by RDKit from the structure and then verified against a parser written independently of the generator. A row survives only if both agree.

That is what makes the released numbers checkable: a reviewer holding these files can recompute every published result on a laptop with no GPU, and disagreeing with us requires disagreeing with RDKit.

Composition

	rows	share
total	90,000
distinct SMILES	76,401
status ok	67,500	75.0%
status invalid_structure	18,000	20.0%
status out_of_scope	4,500	5.0%

A quarter of the corpus is structures that cannot exist or cannot be answered. This is deliberate and it is the part most corpora omit. The base model answers every impossible structure it is shown, all 96 of them in our evaluation, confidently and in the requested format. Knowing when to refuse is a capability, and a corpus made only of answerable questions cannot teach it.

Why a structure is rejected

reason	rows
valence_error	8,323
unbalanced_parenthesis	3,702
element_not_supported	2,935
kekulization_failure	2,477
unknown_element	2,019
size_out_of_range	1,565

Question families

category	rows	what it asks
compute	49,500	counts, paths, formulas, per-atom hydrogen
chem_invalid	10,800	structures that violate chemistry
graph2smiles	10,800	write the SMILES for a described graph
syntax_invalid	7,200	malformed SMILES strings
near_miss	7,200	structures that look valid and are not
out_of_scope	4,500	unsupported elements or sizes

Eight question families are balanced within 3 percent of each other, so no single field dominates the training signal.

Provenance

source	rows	what it is
scaffold_enumerator	75,345	our generator, enumerating ring systems and substituents
wwPDB CCD	14,655	real deposited chemical structures from a public database

The wwPDB Chemical Component Dictionary rows exist so the corpus is not purely our own idiom. They are also the basis of the strongest generalization evidence in the model card: performance on real external molecules is statistically indistinguishable from performance on unseen scaffolds from our own generator, which is not what a memorizing model produces.

Everything is CC0-1.0. The wwPDB CCD is public domain and our generated rows are released the same way.

The v2 continuation shard

A 30,000-row retrain file trains the v2 continuation: a 12,000 row hard shard aimed at two diagnosed defects, plus an 18,000 row seeded random replay of the frozen corpus.

The replay is not optional. Fine-tuning a converged adapter on a narrow hard distribution is the textbook setup for catastrophic forgetting, and the hard shard is 100% polycyclic and kekulization material while rows at 0 to 2 rings are 46% of the held-out slices. The replay is a seeded random sample, so it preserves the original class mix rather than a curated subset.

One consequence worth stating plainly, because it is visible in the data: the continuation mix shifts the reason-code balance. kekulization_failure becomes the most common reason, while valence_error falls from 9.2% of the v1 corpus to 5.7% of the continuation mix. That is intended, since kekulization was the measured weak spot, but it is also the kind of shift that can cost accuracy on a reason code that was already solved, so it is measured rather than assumed.

Verification

• 17 integrity checks pass on the corpus and on every held-out slice.
• 200 of 200 deliberately corrupted rows are rejected by the gates. A guard never proven able to fail is not a guard, which is the most expensive lesson this project has learned.
• Decontamination: 0 of 300 held-out molecules share an InChIKey or a prompt with any of the 90,000 training rows.
• Dual-solver agreement: RDKit and an independently written parser must agree, or the row is dropped.

Held-out slices

Released alongside the corpus, disjoint from it on id, prompt and structure:

slice	rows	what it isolates
mp_held_seen	300	familiar scaffolds, the competition's own yardstick
mp_held_novel	500	ring systems and substituents absent from all 90,000 training rows
mp_real	200	real wwPDB structures, external to our generator
mp_hard	300	4 to 9 rings with closure-critical path queries
mp_para	200	paraphrased prompts, phrasing robustness
mp_random	200	randomized atom ordering, same molecules
mp_smolinstruct	200	a public benchmark, not ours at all

Adaptive Data

Enhancement was applied to instruction sentences only, never to a completion, a computed label, a SMILES string or the CONVENTIONS block, with a gated local merge asserting the STRUCTURE and CONVENTIONS blocks are byte-identical before and after. A rewritten instruction cannot corrupt a verified answer.

Where a pass rewrote a value the labels depend on, it was aborted and disclosed rather than merged. That decision is documented in docs/adaption/molperceive_canary_verdict.md. Shipping silently corrupted labels to raise a row count would have destroyed the one property this corpus is built on.

Try it, and everything that backs it

Live side by side demo: https://huggingface.co/spaces/Jainamshahhh/molperceive-demo Enter your own input and watch the base model and MolPerceive-4B answer it under identical greedy decoding. The GPU backend scales to zero, so a cold first request takes about a minute.

	Hugging Face	Kaggle
model	https://huggingface.co/Jainamshahhh/molperceive-4b	https://www.kaggle.com/datasets/jainamshahhh/molperceive-4b
dataset	https://huggingface.co/datasets/Jainamshahhh/molperceive	https://www.kaggle.com/datasets/jainamshahhh/molperceive-smiles-structure-qa

Released on both platforms, with the scorer, every eval slice, the per row verdicts and significance.py alongside, so every number on this page can be recomputed rather than trusted.

Limitations

Organic main-group chemistry as written in SMILES: C, H, N, O, S, P, F, Cl, Br, I, B and Si, at 5 to 40 heavy atoms. No stereochemistry, no organometallics, no 3D conformation, no reaction prediction. Atom indices follow RDKit's ordering after MolFromSmiles.