A minimal implementation of Data Maps for the Huggingface libraries from the "Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics" paper.
This project contains callbacks that watch training dynamics
(i.e. how well each gold label could be predicted after each epoch).
This allows the creation of dataset maps as the one below that helps to
find out which samples of a dataset are easier and which are harder to learn.
An example on how to use it is given in the main.py file
A more complete version (i.e. the original implementation) can be found here
and a very good implementation as a tensorflow callback that served as template
for this project can be found here under learning.DataMapCallback