ValidMol

Description

ValidMol is an environment for evaluating an agent's ability to complete and fix corrupted molecular SMILES strings. Given a partial or corrupted SMILES notation, the agent must produce a valid, chemically stable molecule. Tasks apply four corruption strategies to real PubChem compounds: truncation, character deletion, ring opening, and bond corruption.

Capabilities

• Completing and repairing corrupted SMILES molecular representations
• Understanding chemical structure and bonding rules
• Reasoning about molecular stability and validity
• Working with cheminformatics concepts (ring closure, bond types)

Compute Requirements

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

Tasks

There are two splits in this environment:

• train: 782 molecular completion tasks
• test: 132 molecular completion tasks

Tasks are generated from PubChem compounds (CIDs 2500-3500 for train, CIDs 4500-4700 for test) with four corruption types:

Reward Structure

This is a sparse, verifiable reward environment. The agent calls submit_completion with a SMILES string and receives a binary reward based on four-tier RDKit validation:

1. Fidelity: the submission repairs the given partial SMILES only where it is broken — the corruption metadata records where each molecule was damaged, and every other run of the partial must survive verbatim, bond symbols included. A submission that fails this positionally is still accepted if it is the intended molecule, compared canonically.
2. Syntactic validity: RDKit parsing succeeds
3. Structural validity: RDKit sanitization passes
4. Chemical stability: No peroxides, hydrazines, or conjugated alkynes

• 1.0: All four tiers pass
• 0.0: Any tier fails

No LLM graders are used.

Data

Task data is generated from PubChem compounds via the PUG-REST API. Quality filters include SMILES length 10-80, heavy atoms 5-50, and a maximum of 5 rings. Data is stored on the OpenReward platform.

Compounds are also screened against the same stability patterns the grader uses, so a compound the grader would reject never becomes a task. Without that screen a task is unsolvable, since the agent is asked to reconstruct a molecule the grader then refuses.

The 17 tasks that predated the screen and whose reference compound the grader rejects were removed from the corpus, which is why the counts above are 782 and 132.

Tools

Time Horizon

Single-turn. Agents receive a corrupted SMILES string and submit one completion.

Environment Difficulty

[Put environment difficulty statistics here]

Other Environment Requirements

There are no further environment requirements; ValidMol works out of the box with the OpenReward endpoint.

Safety

Agents in ValidMol complete molecular structures from corrupted inputs. There is a dual-use concern in that improved molecular generation capabilities could be applied to both beneficial and harmful purposes.

Citation

@dataset{GRValidMol,
  author    = {General Reasoning Inc. Team},
  title     = {ValidMol},
  year      = {2026},
  publisher = {OpenReward},
  url       = {https://www.openreward.ai/GeneralReasoning/ValidMol}
}