Skip to content

Latest commit

 

History

History
24 lines (15 loc) · 663 Bytes

README.md

File metadata and controls

24 lines (15 loc) · 663 Bytes

hf-sprint-diffusion-lm

Implementing controllabale text generation with diffusion for HuggingFace JAX/Diffusers community sprint.

References:
paper https://arxiv.org/pdf/2205.14217.pdf
code https://github.com/XiangLi1999/Diffusion-LM
minimal implementation https://github.com/madaan/minimal-text-diffusion

ToDo

  • add sprint-related requirements (docstring, push to hub)

  • add evaluation

  • [x ] add inference pipeline

  • add reloading model to continue training

  • add training stopping criteria

  • [ x] add other datasets

  • add different tokenizers (SP)

  • [ x] add other padding modes (only block now)