🧬 HippoCrates

A hyper-sterilized, ready-to-use molecular dataset for fine-tuning Chemical LLMs.

---

💡 Overview

Welcome to HippoCrates, curated by ZemResearch. This dataset is specifically designed to teach Large Language Models (LLMs) the universal language of chemistry: SMILES.

We extracted raw molecular data from the legendary QM9 (for basic organic structures) and ChEMBL (for bioactivity and medicinal characteristics) databases, and transformed them into an instruction-following format perfect for fine-tuning text-generation models.

🧼 The "Double Clean" Sterilization Process

Most raw datasets contain broken molecules that defy the laws of physics. We fixed that.

1. RDKit Validation: Every single SMILES string in this dataset has been run through RDKit. If a molecule violated fundamental chemical bonding laws (e.g., hypervalent carbon or 7-bonded phosphorus), it was deleted.
2. Deduplication: We removed all duplicate rows and empty values.
3. Result: 1,460,203 rows of pure, medically and chemically valid molecular data.

📦 Wait, why is the file only 38 MB?

If you look at the files, you might wonder how 1.46 million rows fit into a tiny ~38 MB file.

This dataset uses the Apache Parquet format with Google's Snappy compression. Because SMILES strings have highly repetitive patterns (like c1ccccc1), Snappy compresses the data aggressively. When you load it into your Python environment, it will instantly decompress into the full 1.46 million rows, saving your storage and speeding up your GPU data-loading time!

💻 How to Use (Quick Start)

You don't need to manually download the files. Just use the datasets library by Hugging Face. It will automatically download, cache, and structure the data for you:

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("ZemResearch/HippoCrates-1.4M-SMILES-Clean")

# Check the total rows
print(f"Total verified molecules: {len(dataset['train'])}")

# Look at the first data point
print(dataset['train'][0])

📝 Data Structure

The dataset is formatted for standard Instruction-Tuning (Alpaca/ChatML style). Each row contains three columns:

• instruction: The prompt given to the AI (in standard global English).
• input: Kept empty ("") for this specific dataset version.
• output: The exact, RDKit-verified SMILES string.

Example Row:

{
  "instruction": "Design a small organic molecular structure (SMILES) that adheres to basic chemical bonding laws.",
  "input": "",
  "output": "CC(=O)Oc1ccccc1C(=O)O"
}

🤝 Citation & Collaboration

Created with ❤️ by ZemResearch. If you use this dataset for your academic papers or to train your own medical AI, we'd love to hear about it! Feel free to open a discussion in the community tab.