tokenizer-scratch / README.md
preciousobi's picture
Upload 2 files
391b4ec verified
|
Raw History Blame Contribute Delete
764 Bytes
metadata
license: cc-by-4.0
tags:
  - approx-gelu
  - concat-mlp
  - constant-warmup
  - contrastive
  - instancenorm
  - kaiming-normal
  - mae
  - rmsprop
  - sparse
  - xlarge

finetune.py

Model Overview

A xlarge-scale implementation of the mae architecture, built for contrastive tasks.

Architecture

  • Architecture: mae
  • Scale: xlarge
  • Attention: sparse
  • Fusion strategy: concat mlp
  • Task head: contrastive
  • Activation: approx gelu
  • Normalization: instancenorm
  • Initialization: kaiming normal

Training

  • Optimizer: rmsprop
  • LR scheduler: constant warmup

Files

  • finetune.py — main artifact of this repository

License

See the license field above.