Ether

Description

Ether is an environment for evaluating chemistry reasoning where all answers are molecules in SMILES format. Based on the ether0-benchmark from FutureHouse, tasks include reaction prediction (USPTO/ORD), molecular captioning (PubChem), GHS classification prediction, IUPAC name to SMILES conversion, and molecular design. Each task category contains approximately 25 questions for balanced evaluation.

Capabilities

• Chemistry question answering with SMILES output
• Reaction prediction and retrosynthesis
• Molecular property reasoning
• IUPAC name to structure conversion
• Molecular design and modification

Compute Requirements

Agents are given a standard environment with no sandbox or file system access.

License

CC BY 4.0

Tasks

There is one split in this environment:

• test: 325 chemistry questions

All answers are molecules represented as SMILES strings. Task types include:

• Reaction prediction
• Molecular captioning
• GHS classification
• IUPAC to SMILES conversion
• Molecular design

Reward Structure

This is a sparse, verifiable reward environment with domain-specific grading:

1. Agent submits an answer containing a SMILES string
2. An LLM (gpt-4o) extracts the SMILES from the response
3. Domain-specific reward functions from the ether0 package evaluate correctness:
   • Exact match: String equality for IUPAC/formula tasks
   • Molecular validity: RDKit parsing and sanitization
   • Structure similarity: Tanimoto similarity for design tasks
   • Reaction correctness: Product matching for synthesis tasks
4. Reward ranges from 0.0 to 1.0

Data

Data is sourced from the futurehouse/ether0-benchmark HuggingFace dataset. Tasks are derived from commonly used chemistry benchmarks including USPTO, ORD, and PubChem.

Tools

Time Horizon

Single-turn. Agents receive a chemistry question and submit one answer.

Environment Difficulty

ether0 (the specialized chemistry model) outperforms general-purpose LLMs on these tasks.

Other Environment Requirements

OpenAI API key required for SMILES extraction. Pass via secrets={"openai_api_key": "..."}.

Safety

Agents in Ether answer chemistry questions and generate molecular structures. There is a dual-use concern as molecular generation capabilities could be applied to both beneficial and harmful purposes.

Citation

@article{narayanan2025ether0,
  title={Training a Scientific Reasoning Model for Chemistry},
  author={Narayanan, Sunil Muralidhar and Braza, James D. and Griffiths, Ryan-Rhys and Bou, Adri{\`a} and Wellawatte, Geemi and Ramos, Michael C. and Mitchener, Lewis and Rodriques, Samuel G. and White, Andrew D.},
  journal={arXiv preprint arXiv:2506.17238},
  year={2025},
  url={https://arxiv.org/abs/2506.17238}
}