A gradient-derived trainable activation, xIELU, lowers language model perplexity versus ReLU2 and SwiGLU at matched parameter counts in 125B-token trainings.
Scal- ing fp8 training to trillion-token llms, 2024
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Deriving Activation Functions Using Integration
A gradient-derived trainable activation, xIELU, lowers language model perplexity versus ReLU2 and SwiGLU at matched parameter counts in 125B-token trainings.