Interactive visualization and exploration of scientific papers from the Aella open science dataset.

This project is a collaboration between Inference.net and LAION. LAION curated the original dataset which is about ~100m scrapped scientific and research articles and Inference.net fine-tuned a custom model to extract structured summaries from the articles. This repo contains a visual explorer for a small subset of the extracted dataset.

View the live explorer at https://aella.inference.net.

A web application for exploring scientific papers with semantic embeddings, dimensionality reduction, and clustering visualizations.

• Frontend: React + TypeScript + Vite with D3.js for interactive visualizations
• Backend: Python FastAPI serving data from SQLite (D1 in production)
• Storage: SQLite locally, Cloudflare D1 + R2 in production

You'll need the following tools installed:

• Python 3.11+ - Download
• uv - Python dependency management - Install
• bun - JavaScript runtime - Install
• Task - Task runner - Install

Install all dependencies:

task setup

This will install both backend and frontend dependencies.

Download the database from R2:

task db:setup

This will download the SQLite database to backend/data/db.sqlite.

Run the backend and frontend in separate terminals:

Backend (Terminal 1):

task backend:dev

Frontend (Terminal 2):

task frontend:dev

The application will be available at:

• Frontend: http://localhost:5173
• API: http://localhost:8787
• API Docs: http://localhost:8787/docs

The code for the data pipeline that we used to construct this dataset is not yet open source, mostly because it was setup for a one-time process and not production-ready.

However, the general process was:

1. Initial data extraction and filtering

• Ran a pipeline to generate the summaries
• Excluded specific non-scientific content and failed summaries
• Compiled results for further processing

2. Semantic Embedding

• Generates 768-dimensional embeddings using SPECTER2 (allenai/specter2_base)
• Processes papers in batches with GPU acceleration support
• Stores embeddings as binary blobs for similarity search

3. Visualization & Clustering

• Reduces embeddings to 2D coordinates using UMAP with cosine distance
• Applies K-Means clustering with automatic optimization (20-60 clusters via silhouette scores)
• Generates initial cluster labels using TF-IDF analysis of titles and fields

4. LLM-Curated Labels

• Applies manually reviewed, domain-specific cluster labels
• Improves interpretability over automated TF-IDF labels

Deploy to Cloudflare:

task deploy

This will prompt you to deploy the backend API and/or frontend.

We welcome contributions to this project! Here's what you should know:

Bug Fixes & Minor Improvements

• Bug fixes are always welcome! Please submit a PR with a clear description of the issue and fix.
• Minor improvements to documentation, code quality, or performance are appreciated.

New Features

• This project is intentionally scoped as a one-time preview of this dataset.
• We are generally not planning to greatly expand the functionality beyond its current scope.
• If you want to add significant new features, we encourage you to fork the project and build on it!

Before Submitting a PR

• Ensure your code passes linting and formatting checks:

  task check
  

• Keep changes focused and well-documented.
• Test your changes with sample data when applicable.

MIT License - see LICENSE file for details.