SciReasoner

OpenCompass suite of scientific sequence and text tasks from the SciReasoner paper, spanning molecules, proteins, materials, and DNA/RNA.

Also known as: SciReasoner eval, SciReason

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.
Categorydomain
Subcategoryscientific translation, extraction, property, and generation tasks across molecules, proteins, materials, DNA/RNA
Page statusactive
Metrictask-dependent (accuracy, ROUGE, MAD/MAE, Spearman, SMILES/IUPAC match, and others)
Directionhigher_is_better
PublisherOpen Science Lab (SciReason)

What it measures

SciReasoner here is the evaluation suite that ships with the SciReasoner foundation-model paper, not the weights. A model is asked to move between natural language and scientific objects: SMILES, formulae, proteins, DNA/RNA, and crystal or molecule properties. The paper groups work into translation, knowledge extraction, property prediction, property classification, and conditional or unconditional generation, covering up to 103 tasks. OpenCompass concatenates named slices (GUE, bio_instruction, LLM4Mat, Mol-Instructions, PEER, OPI, USPTO retrosynthesis, bulk modulus, composition, unconditional generation, and others) into `scireasoner_full_datasets` plus a mini cut of each slice.

Task format

Zero-shot GenInferencer on `{input}` unless a slice adds few-shot ICE (smol/LLM4Chem-style configs). Each slice has its own Dataset class and evaluator (accuracy, ROUGE, MAD/MAE, Spearman, SMILES match, and others). Mini configs set mini_set True. Official eval scripts live in open-sciencelab/SciReason on OpenCompass v0.4.2.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub