README.md

metadata

language:
  - en
license: apache-2.0
task_categories:
  - text-generation
  - text-classification
  - question-answering
tags:
  - chemistry
  - toxicology
  - medical
  - biosafety
  - smiles
  - instruction-tuning
  - drug-discovery
pretty_name: HippoXic - Toxicology & Bio-Safety Instruct
size_categories:
  - 10K<n<100K
dataset_info:
  features:
    - name: instruction
      dtype: string
    - name: input
      dtype: string
    - name: output
      dtype: string
  splits:
    - name: train
      num_bytes: 331500
      num_examples: 10630

☠️ HippoXic

A premium instruction-tuning dataset for teaching LLMs chemical toxicology, clinical safety, and side effects.

---

💡 Overview

Welcome to HippoXic, curated by ZemResearch. If our previous dataset (HippoCrates) taught AI how to generate molecules, this dataset teaches AI how to keep us safe from them.

We combined three legendary bio-informatics databases into one streamlined, chat-ready format:

1. Tox21: For biological toxicity pathways (Does it damage DNA? Does it stress cells?).
2. ClinTox: For FDA approval status and clinical trial failures.
3. SIDER: For real-world clinical side effects on human organs.

This is the perfect base to fine-tune your Llama, Qwen, or DeepSeek models into expert toxicologists!

🧼 The "Strict Sterilization" Process

Quality > Quantity. We started with over 14,000 raw molecules, but we only kept the absolute best:

1. RDKit Validation: We ran every single SMILES string through RDKit. If a molecule was chemically impossible or structurally broken, it was ruthlessly deleted.
2. Brutal Deduplication: To prevent model overfitting and data leakage, we merged and dropped all duplicated molecules across the three source datasets.
3. Result: 10,630 rows of 100% chemically valid, non-repeating instruction data.

📦 Wait, why is the file only ~331 KB?

Don't let the file size fool you! Similar to HippoCrates, this dataset uses the Apache Parquet format paired with Google's Snappy compression.

Because text instructions and SMILES strings share highly repetitive text patterns, Snappy compresses them extremely well. A ~331 KB file is super fast to download, but the moment you load it into your Python script, it instantly unpacks into over 10,000 dense rows of medical knowledge. Save your bandwidth, speed up your GPU!

💻 How to Use (Quick Start)

Using the Hugging Face datasets library, you can load this into your workspace in less than a second:

from datasets import load_dataset

# Load the dataset directly from ZemResearch
dataset = load_dataset("ZemResearch/HippoXic")

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

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

📝 Data Structure

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

• instruction: The human prompt asking the AI to analyze the molecule.
• input: The exact, RDKit-verified SMILES string of the chemical.
• output: The detailed AI response (e.g., FDA safety status, toxic pathways, or side effects).

Example Row (From ClinTox subset):

{
  "instruction": "Evaluate the clinical safety of this compound regarding FDA approval and clinical toxicity risk.",
  "input": "CC1=C(C(C(=C(N1)C)C(=O)OCC2=CC=CC=C2N=C=O)C3=CC(=CC=C3)C(F)(F)F)C(=O)OC",
  "output": "Warning: This compound has a high clinical toxicity risk and failed clinical trials."
}

🤝 Citation & Collaboration

Created with ❤️ by ZemResearch. If you use HippoXic to train your bio-safety AI, research chemical toxicity, or write an academic paper, we'd love to hear about it! Drop by our community tab and say hi.