Skip to content

Latest commit

 

History

History
84 lines (48 loc) · 2.15 KB

todo.md

File metadata and controls

84 lines (48 loc) · 2.15 KB

Milestones

(description)

  • (list of things we need to get done)

Status: (status)

Blockers: (list of things we have to de before we can do this milestone)

Synthetic Identity Initialization

Handcraft weights of T5 model to perform the identity task out of the box.

  • identified a bug that blocks benchmarks on this model

Status: complete but buggy

Synthetic Set Initialization

Handcraft weights of T5 model to perform the set deduplication task out of the box.

  • needs layer to replace nonduplicate characters with <pad>

Status: mostly complete

Finetune benchmarks

identity (pretrained)

  • Use results from paper to compare to downstream performance of the synthetic initialized identity.

status: complete (from paper)

set (pretrained)

  • Use results from paper to compare to downstream performance of the synthetic initialized set.

status: complete (from paper)

baseline (random initialization)

  • Run finetuning on model with random initializations.
  • Build a script for finetuning that can take an arbitrary model and return a model optimized for a given downstream task.
  • Get training results and compile them to a graph.
  • status: not started
  • needs finetuning script for huggingface transformers and downstream tasks

blockers: none (this is ready and needs to be done soon)

identity (initialized)

  • Run finetuning on model with identity synthetic initialization.
  • Use the finetuning script from (baseline).
  • Get training results and compile them to a graph.

status: not started

blockers:

  • baseline
  • bug fix for decoder

set (initialized)

status: not started

blockers:

  • baseline
  • synthetic nonduplicate padding layer

Generalized Synthetic Weight Generation

status: todo, fingers crossed

Slides for Final Presentation

status: todo, very important!

  • we have been painting very localized pictures in our two demos
  • for our final demo we need to clearly restate our goals and our line of thinking
  • then we need to ideally show the results
  • as a stretch goal it would be nice to demostrate the generalized synthetic weight generation