Oduwo/drugbank-clean

DrugBank Clean — Instruction-Tuning Dataset for Pharmacology AI

A high-quality, instruction-tuned dataset of 45,805 question-answer pairs derived from DrugBank, designed for fine-tuning LLMs on pharmacological knowledge. Built to power the next generation of clinical AI systems with deep pharmacology expertise.

🎯 Why This Dataset

Most general-purpose LLMs lack deep pharmacological knowledge — they hallucinate dosages, miss drug interactions, and oversimplify clinical mechanisms. This dataset addresses that gap by providing clean, structured, instruction-tuned data specifically for fine-tuning models on:

• 💊 Drug overviews and clinical descriptions
• 🩺 Indications and clinical use cases
• 🧬 Mechanism of action (molecular level)
• 📊 Pharmacodynamics & pharmacokinetics
• ⚠️ Drug-drug interactions
• 🔬 Absorption, half-life, protein binding
• 🚨 Toxicity profiles and overdose risks

📊 Dataset Statistics

Metric	Value
Rows	45,805
Format	JSONL (instruction-output pairs)
Fields	instruction, output
Source	DrugBank database (cleaned & restructured)
License	MIT
Language	English

🧬 Question Categories Covered

The dataset covers 10+ pharmacology dimensions per drug:

• ✅ Drug overviews ("Give me an overview of [drug]")
• ✅ Clinical indications ("What are the clinical indications for [drug]?")
• ✅ Prescriber guidance ("When should a clinician prescribe [drug]?")
• ✅ Mechanism of action (high-level + molecular)
• ✅ Pharmacodynamics
• ✅ Absorption profiles
• ✅ Half-life data
• ✅ Protein binding percentages
• ✅ Elimination pathways
• ✅ Toxicity & overdose risks
• ✅ Drug-drug interaction warnings

🚀 Use Cases

This dataset is ideal for:

• Fine-tuning LLMs for clinical decision support (LoRA, QLoRA on Llama, Mistral, BioMistral, Phi)
• Training pharmacology chatbots (patient-facing or clinician-facing)
• Retrieval-Augmented Generation (RAG) systems for healthcare
• Drug information question-answering systems
• Educational AI tutors for pharmacy students
• Research benchmarks for medical LLM evaluation

📥 Quick Start

Load with datasets library:

from datasets import load_dataset

dataset = load_dataset("Oduwo/drugbank-clean")
print(dataset)
print(dataset["train"][0])

Sample row format:

{
  "instruction": "Give me an overview of the drug Lepirudin.",
  "output": "Lepirudin is a recombinant hirudin formed by 65 amino acids that acts as a highly specific and direct thrombin inhibitor..."
}

Fine-tune with QLoRA (example):

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import LoraConfig, get_peft_model
from datasets import load_dataset

dataset = load_dataset("Oduwo/drugbank-clean", split="train")
model = AutoModelForCausalLM.from_pretrained("BioMistral/BioMistral-7B")
# ... continue with your fine-tuning pipeline

🛠️ Data Pipeline

The dataset was built using a custom streaming XML parser for the DrugBank database, with the following processing steps:

1. Extraction — DrugBank XML parsed via streaming for memory efficiency
2. Restructuring — Raw drug records transformed into instruction-output pairs
3. Deduplication — Removed exact and near-duplicate entries
4. Format standardization — Converted to clean JSONL for LLM training
5. Quality validation — Manual review of sample entries

⚠️ Limitations

• Source: Derived from DrugBank — inherits any limitations of that source
• Language: English only
• Currency: Reflects DrugBank version at time of curation
• Use case: Designed for AI training; not a substitute for clinical references like UpToDate, Lexicomp, or Micromedex
• Clinical disclaimer: Models trained on this data are for educational and research purposes only. Always verify with authoritative clinical sources before applying in patient care.

📚 Citation

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

@dataset{oduwo2026drugbank_clean,
  author = {Emmanuel Bain Oduwo},
  title = {DrugBank Clean: Instruction-Tuning Dataset for Pharmacology AI},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Oduwo/drugbank-clean}
}

🔗 Related Projects

This dataset powers several of my healthcare AI projects:

• Kemirix — Self-hosted pharmacology LLM (BioMistral 7B + QLoRA)
• DrugD Clinical AI — Drug information & interaction platform
• PharmaCliniq Pro — Clinical decision support tool
• Africa Pharmacovigilance ADR Dataset — Open-source African drug safety data

👨‍💻 Author

Emmanuel Bain Oduwo
Healthcare AI/ML Engineer | B.Pharm Student | Founder, TechFryz
Building production-grade clinical AI tools that bridge pharmacology and machine learning, with a focus on African healthcare contexts.

• 📧 Email: bainemmanuel86@gmail.com
• 📱 WhatsApp: +254 707 862 452

🤝 Contributing

Found an issue? Have suggestions? Open a discussion on the Community tab. Contributions to expand coverage (e.g., African pharmacology, traditional medicine integrations) are especially welcome.

📜 License

MIT License — free for research, educational, and commercial use with attribution.