Inference · Sampling Real scores from the model you trained

Calm, or
unhinged.

The model doesn't pick the next character — it scores every character, then rolls a weighted die. Temperature is how loaded the die is. Turn it down and it always bets the safe favourite; turn it up and it starts taking wild chances. Same model, same scores — one dial between sensible and deranged. This is the dial you were promised at the very first station, where the model simply took its top bet every time and repeated itself for it: that was this dial, held at its calmest setting.

Its odds on the next character — the 12 likeliest of the 65 it knows. ␣ is a space, ⏎ a new line.

Temperature0.80

Top-k

Twenty separate rolls for the same next character — not twenty characters of writing.

This is where "creativity" comes from. There's no magic — just a probability list and a die. Low temperature is why a model can sound repetitive and safe; high temperature is why it can be surprising, or spew nonsense. top-k is a guardrail: only let the die land on the k most likely characters, so it can loosen up without ever picking something absurd.

uv run python 04_pretrain/sample_extract.py # reads real next-token logits from the trained model

Real scores from the trained model · this page divides them by the temperature, crosses out everything outside your top-k, then turns what is left into percentages — the same three steps the model takes each time it writes a character · nothing faked. Built by AgentSure.