💊 Hippolv

A highly curated, instruction-tuning dataset for teaching AI how drugs behave in the human body (ADMET & Solubility).

---

💡 Overview

Welcome to Hippolv, the third specialized dataset curated by ZemResearch. While our previous datasets focused on general molecular structures and toxicology, Hippolv is engineered to bridge the gap between pure chemistry and clinical pharmacology.

This dataset focuses on ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) and aqueous solubility. We sourced raw, high-quality clinical experimental data from the prestigious OpenADMET Challenges (expansionrx and pxr-challenge), and transformed them into a clean, instruction-following format. It's the perfect fuel to fine-tune your Chemical LLMs into expert pharmaceutical assistants.

🧼 The "Deep Clean" Sterilization Process

High-quality models require high-quality data. We didn't just scrape this data; we rigorously sterilized it using a strict RDKit pipeline:

1. Chemical Validation: Every molecule was passed through RDKit. If the SMILES string was broken or chemically invalid, it was instantly dropped.
2. Canonicalization: All valid molecules were converted into their standard Canonical SMILES (with isomeric properties preserved) to ensure structural consistency.
3. Deduplication: We aggressively dropped duplicates based on their true chemical structure, not just string similarity. We also removed any rows with missing ADMET values.
4. Result: A hyper-sterilized dataset of pure, instruction-ready ADMET data, free from data leakage and benchmark conflicts.

📦 Why is the file so tiny?

If you check the files, you might be surprised to see that thousands of rows of pharmaceutical data fit into a file of around 1 MB.

This is the magic of the Apache Parquet format paired with Google's Snappy compression. Since SMILES strings and repetitive text formats compress incredibly well, you get a feather-light file to download. Once loaded into your Python environment, it instantly decompresses to its full size, saving your bandwidth and keeping your GPU data-loading ultra-fast!

💻 How to Use (Quick Start)

You don't need to manually download anything. Just use the datasets library by Hugging Face to load it directly into your workflow:

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("ZemResearch/Hippolv")

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

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

📝 Data Structure

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

• instruction: The prompt asking the AI to analyze the molecule's properties.
• input: The verified Canonical SMILES string of the molecule.
• output: A structured narrative detailing the compound's ADMET and solubility metrics.

Example Row:

{
  "instruction": "Analyze the aqueous solubility and ADMET profile for the following chemical compound.",
  "input": "CC(=O)Nc1ccc(O)cc1",
  "output": "ADMET Metrics:\n- LogD: 0.46\n- Kinetic Solubility (KSOL): High\n- HLM CLint: 12.5\n- Caco-2 Permeability: Good"
}

(Note: The example output reflects the combined narrative format generated during our preprocessing pipeline).

🤝 Citation & Collaboration

Created with ❤️ by ZemResearch. If you use Hippolv for your academic research, drug discovery pipelines, or to train your own AI models, we'd love to hear about it! Feel free to open a discussion in the community tab.