Skip to content

Language modelling task to infer shapes of matrices - One of the most difficult tasks for models like GPT-3, GPT-J

License

Notifications You must be signed in to change notification settings

Muennighoff/matrixshapes

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

27 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

matrixshapes

This task primes language models to keep track of matrixshapes.

To generate a dataset, run: git clone https://github.com/Muennighoff/matrixshapes.git & cd into the repository.
Then run python generate.py for an example datum. Feel free to change the hyperparameters, such as more operations or higher dimensions (both will need more RAM).

To create a json dataset run python create_json.py --num 1000 --cont 0.5 for 1000 examples where a maximum of 500 shapes is also contained in the operations.

See the task.json for a dataset with num 5000, cont 0.5 & default values for generate.

This task is part of Google's BIG-Bench, see here.

@misc{matrixshapes,
  author = {Muennighoff, Niklas},
  title = {{Keeping track of matrix shapes after transformations}},
  howpublished = {\url{https://github.com/Muennighoff/matrixshapes}},
  year = 2021,
  month = February
}

About

Language modelling task to infer shapes of matrices - One of the most difficult tasks for models like GPT-3, GPT-J

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages