Glossario IA

📖

termini

Causal Masking

Technique in the decoder that masks all future positions to ensure that the prediction for position i only depends on positions 1 to i, respecting the auto-regressive nature of generation.

📖

termini

Output Projection

Final linear layer that projects the decoder representations into the vocabulary space, followed by a softmax to produce a probability distribution over possible tokens at each output position.