urchade/GLiNER

Clarification of model size in paper

Closed this issue · 2 comments

Hi, I noticed that in paper the GLiNER s/m/l are of size 50M, 90M, 0.3B respectively, while the published pretrained models have size of 166M, 209M, 459M. Could you please clarify what makes this difference?

Also, why GoLLIE is excluded from the evaluation of the 20 NER dataset in the paper?

Hi @SCccc21

  1. All model sizes exclude the embedding layer parameters
  2. not sure if in gollie paper, the authors evaluated on all these datasets

Do not hesitate if you have any other questions :)