Skip to contentExperiment
From context to the next token
Separate the model’s probabilities from the rule that selects an output.
Essentials
How it works
A token may be a word, part of a word or a symbol. An autoregressive language model computes scores from context; normalisation produces a distribution. Decoding then selects a token, which joins the context for the next step.
Takeaway
Probability expresses model preference in this context, not the truth of a sentence. Temperature changes sampling without adding knowledge. Our example has only three fictional possibilities.
From concept to practicePut it into practice · ChatGPT
Scientific sources
- Vaswani et al. · 2017Attention Is All You NeedarXiv v7 · 2023Original publication
Teaching synthesis of the cited sources. Published results remain tied to their tasks, protocols and budgets.