README.md

---
title: ZemResearch
emoji: 🧪
colorFrom: blue
colorTo: purple
sdk: static
pinned: false
---

<div align="center">
  <h1>🧬 Hello from ZemResearch!</h1>
  <p><i>Mixing Artificial Intelligence with Chemistry, one dataset at a time.</i></p>
  <hr>
</div>

### 👋 Who We Are
Welcome to **ZemResearch**! We are an open-source research initiative passionate about bridging the gap between computer science and molecular biology. We believe that training specialized, lightweight Large Language Models (LLMs) shouldn't require massive corporate budgets—it just needs incredibly clean data and smart engineering.

### 🎯 What We Do
* **🧹 Extreme Data Cleaning:** We don't just scrape data; we sterilize it. We heavily rely on tools like RDKit to ensure our molecular datasets obey the fundamental laws of chemistry.
* **🤖 Lightweight AI Models:** We focus on fine-tuning accessible, efficient LLMs that can run smoothly without needing massive GPU clusters.
* **🌍 Open Science:** Everything we build is dedicated to the global open-source community. Let's democratize AI drug discovery together!

### 🦛 The Hippo Ecosystem
Our datasets are designed to work together, covering the full drug discovery pipeline — from raw molecular structure all the way to clinical safety. Each one is independently useful, but together they form a complete picture:

**HippoCrates → HippoSynth → HippoTarget → HippoLv → HippoXic**
*(what it is → how it's made → what it binds to → how it behaves in the body → whether it's safe)*

| Dataset | Focus | Size |
|---|---|---|
| 🧬 [HippoCrates](https://huggingface.co/datasets/ZemResearch/HippoCrates-1.4M-SMILES-Clean) | Molecular structures & SMILES | 1.46M rows |
| ⚗️ [HippoSynth](https://huggingface.co/datasets/ZemResearch/HippoSynth) | Chemical reactions & synthesis | 50K rows |
| 🎯 [HippoTarget](https://huggingface.co/datasets/ZemResearch/HippoTarget) | Drug-target interaction | 15.5K rows |
| 🫀 [HippoLv](https://huggingface.co/datasets/ZemResearch/HippoLv) | ADMET & drug behavior in the body | 9.4K rows |
| ☠️ [HippoXic](https://huggingface.co/datasets/ZemResearch/HippoXic) | Toxicology & clinical safety | 10.6K rows |

Every dataset in the ecosystem follows the same curation standard: RDKit-validated, deduplicated, formatted for instruction-tuning (Alpaca style), and compressed with Snappy for a tiny footprint.

### 🚀 Dataset Details

* **[HippoCrates](https://huggingface.co/datasets/ZemResearch/HippoCrates-1.4M-SMILES-Clean):** A massive, heavily sterilized dataset containing 1.46 million molecular structures. It's ready-to-use (in Apache Parquet format) for text-generation and chemical bioactivity fine-tuning.

* **[HippoSynth](https://huggingface.co/datasets/ZemResearch/HippoSynth):** A curated dataset of 50,000 chemical synthesis and reaction examples — covering forward synthesis, retrosynthesis, and lab procedure interpretation. Built from USPTO patent reactions, the Open Reaction Database, and curated chemistry Q&A sources.

* **[HippoTarget](https://huggingface.co/datasets/ZemResearch/HippoTarget):** A drug-target interaction dataset with 15,520 rows, combining real experimental binding data with FDA-approved drug-target pairs. Teaches models which molecules bind to which proteins.

* **[HippoXic](https://huggingface.co/datasets/ZemResearch/HippoXic):** A premium, domain-specific instruction-tuning dataset containing 10,630 highly curated rows focused on chemical toxicology, FDA clinical safety, and real-world side effects. It bridges the gap between molecular structures and clinical bio-safety reasoning.

* **[HippoLv](https://huggingface.co/datasets/ZemResearch/HippoLv):** A hyper-sterilized, feather-light dataset containing 9,465 RDKit-verified rows focused on drug behavior inside the human body (ADMET properties & aqueous solubility). It serves as the ultimate fuel for turning general LLMs into clinical pharmaceutical experts.

### 🤝 Let's Collaborate
Got a cool idea for molecular LLMs, or just want to chat about AI in healthcare? Feel free to explore our datasets, open a discussion in our repositories, or reach out. We are always open to new collaborations!

<br>
<div align="center">
  <i>Stay curious. Keep building.</i> 🚀
</div>