Logits
The raw, unnormalised scores a model produces before softmax.
Overview
Logits are the vector of scores the final layer emits for each vocabulary token, before any normalisation. They solve nothing; they are the model's output in its most basic form. They exist as the place every decoding decision is made: temperature scales them, top-p filters them, and constrained decoding restricts which of them may be chosen.
Technology Graph
How Logits connects to the rest of the stack.
Used by
- Beam Search technique Techniques Logits used by Beam Search High ✓ source
- Constrained Decoding technique Techniques Logits used by Constrained Decoding High ✓ source
- Temperature technique Techniques Logits used by Temperature High ✓ source
- Top-p Sampling technique Techniques Logits used by Top-p Sampling High ✓ source