Perplexity & Generation Metrics - Interactive Visualization

Perplexity measures how 'surprised' a language model is by text - lower is better. BLEU and ROUGE count n-gram overlaps between generated and reference text. BERTScore uses contextual embeddings to measure semantic similarity. This demo shows each metric token-by-token on real generated text.

• Per-token perplexity heatmap - see which tokens the model found surprising vs predictable in a given sequence
• BLEU score breakdown - watch precision at n-gram orders 1, 2, 3, 4 and the brevity penalty computed live
• ROUGE-1, ROUGE-2, ROUGE-L - see recall-oriented overlap scores and where BLEU and ROUGE disagree
• BERTScore - visualize the token-level cosine similarity matrix between generated and reference text
• Understand why low perplexity does not guarantee high BLEU - and why all four metrics can fail on the same text

Part of the EngineersOfAI Interactive 3D - free interactive visualizations covering every major concept in machine learning and AI engineering. Hover any element for a plain-English explanation. No code required.