Skip to content

segyges/pythia-embedding-analysis

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

28 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

What Is Learned, When? | EleutherAI Community Project

reference thesis: https://nsaphra.net/uploads/thesis.pdf

older work (e.g. that thesis) uses lstm, is unclear if any of that transfers to a modern architecture, pythia checkpoints give a good reference point

Goals for the project

As per JDC has mentioned:

  1. Establish things that pythia-12b embeddings have learned when fully trained.
  2. Look through the checkpoints to see at what point the model learns those things/how quickly it learns those things/what the learning curve looks like for all those things the fully trained model learns
  3. See how this extends to other pythia models/sizes.

TODO List:

  • Upload all(or a meaningful subset, e.g. every power of 2) pythia-12b checkpoints to HF
  • Analysis of token meanings/categories in the fully trained pythia-12b model
  • Analysis of what meanings show up when in training pythia-12b
  • Potentially expand this to other pythia model sizes to see if this is true across scales?

Links to data

GSON has uploaded weights here: https://huggingface.co/amphora/pythia-12b-weights

And data on their similarities here: https://huggingface.co/amphora/pythia-12b-weights/tree/main/cos_sim

References

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published