Type a prompt, pick a small language model, run one real forward pass — entirely
in your browser, via transformers.js (WASM/ONNX), no server
inference — and see the actual probability distribution the model puts over
every possible next token. Prompted by a thread asking whether next-token
logits get "absurdly skewed" when a model is mid-way through repeating text
verbatim. Now you can just check.
top token
—
entropy
—
bits over the full vocab
effective choices
—
2^entropy — a "how many tokens feel live" count
Click any token to append it to the prompt and step forward one token — a quick way to watch the distribution stay skewed (or not) across a run.