Talip7/drug-reviews-cleaned

🧪 Drugs.com Cleaned Reviews Dataset

This dataset is a cleaned and structured version of the Drugs.com user reviews dataset, prepared for modern NLP tasks such as text classification and abstractive summarization.

The primary goal of this dataset is to enable reproducible experimentation with user-generated medical text using the 🤗 Hugging Face ecosystem.

---

📊 Dataset Overview

Each entry corresponds to a user-written review of a drug, along with metadata describing the medication, reported condition, and user feedback signals.

Main fields

• patient_id: Unique identifier of the review
• drugName: Name of the drug
• condition: Medical condition treated (lowercased)
• review: Free-text user review (HTML-cleaned)
• rating: User rating (1–10)
• usefulCount: Number of users who found the review helpful
• review_length: Number of tokens in the review

---

🧹 Preprocessing Steps

The raw dataset was processed with the following steps:

• Removal of reviews with missing conditions
• Lowercasing of condition names
• HTML entity cleanup (e.g. &, ')
• Removal of very short reviews (review_length <= 30)
• Deterministic train/validation split for reproducibility

These steps were designed to minimize noise while preserving the semantic richness of user-written reviews.

---

📦 Dataset Splits

Split	Examples
Train	96,959
Validation	41,555
Test	46,108

---

🧠 Supported Tasks & Use Cases

1️⃣ Multi-class Text Classification

The dataset can be used to train classifiers that predict the patient condition based solely on the review text.
This is a challenging real-world setting with 771 distinct conditions and strong class imbalance.

2️⃣ Abstractive Summarization

In addition to classification, the dataset was used with a pretrained summarization pipeline (BART-large-CNN) to generate concise summaries of long user reviews.

This allows exploration of:

• Opinion summarization
• Side-effect abstraction
• Noise robustness in medical text

The cleaned Drugs.com dataset was published to the Hugging Face Hub to ensure reproducibility and ease of reuse. In addition to multi-class classification, an abstractive summarization pipeline (BART-large-CNN) was applied to generate concise summaries of user-written drug reviews.

---

🚀 Example Usage

from datasets import load_dataset

dataset = load_dataset("Talip7/drug-reviews-cleaned", split="test")
print(dataset[0]["review"])

from transformers import pipeline
import html

summarizer = pipeline("summarization", model="facebook/bart-large-cnn")

text = html.unescape(dataset[0]["review"])
summary = summarizer(text, max_length=60, min_length=25, do_sample=False)

print(summary[0]["summary_text"])

---

⚠️ Limitations

• Highly imbalanced class distribution across medical conditions

• User-generated text may contain ambiguity and subjective language

• Summaries may omit fine-grained medical details due to the abstractive nature of the model

• Not intended for medical advice or diagnosis

---

👤 Author

Created and curated by Talip as part of hands-on NLP and LLM experimentation with Hugging Face.

---

📚 Original Data Source

• Drugs.com User Reviews dataset https://archive.ics.uci.edu/ml/datasets/Drug+Review+Dataset+(Drugs.com)

---