Token-by-token generation — deepseek-r1:8b

Before the Script

The 560 tokens generated before the model wrote any code for “create a batch file (minimal) to map a network drive.” Every branch is a token that received nonzero probability at that step; copper is the one the sampler drew.

prompt

token span Ollama's API labels message.reasoning

Labeled “reasoning” because that is the literal field Ollama's API returns (message.reasoning) for this model family, alongside the literal <think>/</think> tokens the model was trained to emit around this span. The label names which tokens the final answer is conditioned on — it is not a claim about cognition. Every token on this page, inside that span or outside it, is produced the same way: softmax over logits, then a weighted random draw.

sampled — was also the top-probability token sampled — fell outside the shown top-k despite lower probability nonzero probability, not sampled

Built with vis-network from a real trace via Ollama's logprobs API: at every step the model's top candidate tokens and their real probabilities were captured, and the next token drawn by the server's own sampling -- not scripted.