What you're looking at

A real sentence-transformer, all-MiniLM-L6-v2, runs entirely in your browser on WebGPU (via transformers.js). It reads thousands of real tweets from the public dair-ai/emotion dataset and turns each one into a 384-dimensional vector that captures its meaning.

Those vectors are projected to 3D with principal component analysis, so sentences with similar meaning sit close together. Colour is each tweet's real emotion label. When you drop in your own sentence, it's embedded live and placed by the same projection, with lines to its true nearest neighbours (cosine similarity) and a k-NN emotion vote.

Honest notes: positions are a PCA shadow of 384-D space, so some structure is lost; the "prediction" is a nearest-neighbour vote, not a trained classifier. Nothing is precomputed or synthetic, the model and every tweet load live.