summykai/chembench-rlvr-test

ChemBench-RLVR: Comprehensive Chemistry Dataset for Reinforcement Learning from Verifiable Rewards

Dataset Description

ChemBench-RLVR is a high-quality, balanced dataset containing 7,001 question-answer pairs across 14 chemistry task types. This dataset is specifically designed for training language models using Reinforcement Learning from Verifiable Rewards (RLVR), where all answers are computationally verifiable using established cheminformatics tools.

Key Features

• 🧪 7,001 balanced QA pairs across 14 chemistry domains
• 🔬 100% local calculations - no external API dependencies
• ⚖️ Perfect task balance - each task has exactly 500 samples
• 🎯 Verifiable answers - all responses computed using RDKit, spyrmsd, and other reliable tools
• 📚 Template diversity - 3 prompt variations per task
• 🌐 Molecular diversity - sourced from 10,000 PubChem compounds
• 📦 Multiple formats - Available in both Parquet and JSONL formats

Dataset Statistics

Overview

• Total Samples: 7,001
• Training Split: 6,300 samples (90%)
• Test Split: 701 samples (10%)
• Generation Time: 2685.7 seconds
• Average per Task: 500 samples
• Zero Duplicates: All QA pairs are unique
• Reproducible: Fixed seed (42) for consistent results

Molecular Complexity Statistics

• SMILES Length: 41.2 ± 16.7 characters (avg ± std)
• Min SMILES Length: 7 characters
• Max SMILES Length: 84 characters
• Median SMILES Length: 45 characters

Text Length Statistics

Questions

• Average Length: 111 ± 20 characters
• Range: 45-160 characters
• Median: 112 characters

Answers

• Average Length: 1683 ± 4695 characters
• Range: 5-24829 characters
• Median: 105 characters

Task Distribution

Complete Task Breakdown

• Bioactivity Prediction: 384 samples- Drug Likeness Assessment: 442 samples- Functional Group Identification: 408 samples- Ghs Hazard Statement Identification: 249 samples- Ghs Pictogram Identification: 248 samples- Hydrogen Bond Properties: 404 samples- Iupac Name Generation: 714 samples- Logp Calculation: 27 samples- Molecular Weight Calculation: 1,255 samples- Molecule Visualization: 717 samples- Reactivity Prediction: 405 samples- Solubility Prediction: 667 samples- Stereochemistry Analysis: 750 samples- Synthetic Accessibility: 331 samples

Chemistry Task Categories

🧪 Core Molecular Properties (6 tasks)

• Molecular Weight Calculation: Exact molecular mass computation using RDKit
• LogP Calculation: Octanol-water partition coefficient prediction
• Aromatic Ring Count: Identification of aromatic ring systems
• Hydrogen Bond Properties: Count of donors and acceptors
• IUPAC Name Generation: Systematic nomenclature from structure
• Molecule Visualization: 2D structural diagram generation

🔬 Advanced Spectroscopy & Structure (3 tasks)

• NMR Signal Prediction: 1H and 13C chemical shift estimation via RDKit fallback methods
• Point Group Determination: Molecular symmetry analysis using RDKit/spyrmsd
• Stereochemistry Analysis: Chiral center identification and stereoisomer enumeration
• Functional Group Identification: SMARTS-based substructure recognition

⚠️ Safety & Hazard Assessment (2 tasks)

• GHS Pictogram Identification: Hazard symbol classification from structure
• GHS Hazard Statement Identification: H-code assignment using chemical patterns

💊 Pharmaceutical Chemistry (4 tasks)

• Drug-Likeness Assessment: Lipinski's Rule of Five evaluation
• Solubility Prediction: Aqueous solubility estimation via group contribution
• Bioactivity Prediction: Pharmacological class prediction from structural features
• Stereochemistry Analysis: Chiral center identification and stereoisomer counting

⚗️ Synthetic Chemistry (2 tasks)

• Synthetic Accessibility: Complexity scoring for synthesis planning
• Reactivity Prediction: Reactive site identification and charge analysis

Task Distribution

• Bioactivity Prediction: 384 samples
• Drug Likeness Assessment: 442 samples
• Functional Group Identification: 408 samples
• Ghs Hazard Statement Identification: 249 samples
• Ghs Pictogram Identification: 248 samples
• Hydrogen Bond Properties: 404 samples
• Iupac Name Generation: 714 samples
• Logp Calculation: 27 samples
• Molecular Weight Calculation: 1,255 samples
• Molecule Visualization: 717 samples
• Reactivity Prediction: 405 samples
• Solubility Prediction: 667 samples
• Stereochemistry Analysis: 750 samples
• Synthetic Accessibility: 331 samples

Dataset Structure

Each sample contains:

• messages: List of conversation turns (user question, assistant answer)
• task: Chemistry task category
• smiles: SMILES string of the molecule
• difficulty: Task difficulty level (easy/medium/hard)

Example Sample

{
  "messages": [
    {
      "role": "user", 
      "content": "What is the molecular weight of the compound with SMILES 'CCO'?"
    },
    {
      "role": "assistant",
      "content": "The molecular weight of ethanol (CCO) is 46.07 g/mol."
    }
  ],
  "task": "Molecular_Weight_Calculation",
  "smiles": "CCO",
  "difficulty": "easy"
}

Computational Methods

All answers are computed using established cheminformatics libraries:

• RDKit: Molecular property calculations, structure analysis
• spyrmsd: Symmetry-corrected molecular analysis
• MDAnalysis: Molecular dynamics and structure processing
• PyTorch: Neural network components (when available)

Usage

Loading the Dataset

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("summykai/chembench-rlvr-test")

# Load specific split
train_data = load_dataset("summykai/chembench-rlvr-test", split="train")
test_data = load_dataset("summykai/chembench-rlvr-test", split="test")

RLVR Training

This dataset is optimized for Reinforcement Learning from Verifiable Rewards:

# Example: Verify molecular weight calculation
from rdkit import Chem
from rdkit.Chem import Descriptors

def verify_molecular_weight(smiles: str, predicted_mw: float) -> bool:
    mol = Chem.MolFromSmiles(smiles)
    if mol is None:
        return False
    actual_mw = Descriptors.MolWt(mol)
    return abs(actual_mw - predicted_mw) < 0.1

Citation

If you use this dataset in your research, please cite:

@dataset{chembench_rlvr_2025,
  title={ChemBench-RLVR: Comprehensive Chemistry Dataset for Reinforcement Learning from Verifiable Rewards},
  author={ChemBench Team},
  year={2025},
  url={https://huggingface.co/datasets/summykai/chembench-rlvr-test},
  note={Generated using RDKit, spyrmsd, and other open-source cheminformatics tools}
}

License

This dataset is released under the MIT License. See LICENSE file for details.

Dataset Generation

• Generated on: 2025-07-29 22:29:32 UTC
• Version: 8.6-post8
• Seed: 42 (for reproducibility)
• Source molecules: PubChem compound database

Acknowledgments

This dataset was generated using:

• RDKit - Cheminformatics toolkit
• spyrmsd - Symmetry-corrected RMSD calculations
• PubChem - Chemical compound database
• Hugging Face - Dataset hosting and distribution