ChemBERTa-2
Chemical foundation model: RoBERTa pretrained on 77M molecules.
About
ChemBERTa-2 is a RoBERTa-style transformer pretrained on 77M unique SMILES from PubChem. It comes in two pretraining flavors — masked language modeling and multi-task regression over 200 RDKit-computed molecular properties — and in 5M, 10M, 77M, and 100M parameter sizes.
It is a representation model, not a predictor out of the box: you fine-tune it (or use its embeddings) for property prediction, ADMET, and similar tasks, where it shows competitive MoleculeNet performance that scales with pretraining data size. Weights are on Hugging Face under the DeepChem org (Apache 2.0), and it runs anywhere Transformers runs.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Competitive downstream performance on MoleculeNet; performance scales with pretraining dataset size (5M → 77M). No single headline score — it is a backbone for fine-tuning, not a task model.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: standard chemical representation backbone; widely downloaded on Hugging Face; mature project; slower recent development; Apache 2.0; multiple sizes freely available. Manual review: a backbone, not a predictor — needs fine-tuning for any real task. License: Apache 2.0 (DeepChem project)..
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · standard chemical representation backbone; widely downloaded on Hugging Face
- · mature project; slower recent development
- · Apache 2.0; multiple sizes freely available
- · manual review: a backbone, not a predictor — needs fine-tuning for any real task