SMolInstruct

A 14-task small-molecule chemistry instruction set of about 3.3 million samples, used both to train LlaSMol and as an OpenCompass evaluation.

Also known as: SmolInstruct, SMolInstruct, LlaSMol dataset

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.

Category	domain
Subcategory	small-molecule chemistry instruction tuning and evaluation (14 tasks)
Page status	active
Metric	task-specific: exact or element match, accuracy, RMSE, METEOR, Morgan Tanimoto
Direction	higher_is_better
Unit	task-specific
Dataset licence	CC-BY-4.0
Publisher	Ohio State University NLP Group

What it measures

SMolInstruct evaluates instruction-following on small-molecule chemistry. OpenCompass scores 14 English tasks: four name conversions (IUPAC and SMILES to formula or SMILES/IUPAC), six property predictions (ESOL, Lipo, BBBP, ClinTox, HIV, SIDER), molecule captioning, molecule generation, forward synthesis, and retrosynthesis. Inputs and targets use tagged SMILES, IUPAC, formulas, numbers, or yes/no. The same collection is the instruction-tuning set for LlaSMol.

Task format

OpenCompass generation. Default smolinstruct_gen.py uses FixKRetriever with one validation example (1-shot) and a chemistry system hint per task. Separate 0-shot instruct configs exist. Tags such as <SMILES> wrap core answers. mini_set can cap each test task at 500 items.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub