Released! Feel free to checkout our model checkpoints here. From our point of view, researchers might be more familiar with using huggingface
, so those checkpoints were in huggingface format.
In each files under directory checkpoint_paths
, we provide the paths to all .tar
files. One could modify the files accordingly and download with commands like one below (For more options, see here).
xargs -n 1 -I{} wget https://arkdata.cs.washington.edu/probe-across-time/{} --no-check-certificate < <(cat path/to/modified/file)
- As an illustrative example, folder
NEWS
denotes the checkpoints from pretraining RoBERTa-base on downsampled RealNews, see more details in the paper. - Within folders like,
NEWS
, you will seeroberta_bz256_savesteps
orroberta_bz256_saveepochs
or both; those names denote how the checkpoints (within those folders) are saved during prertaining by epoch or step. The name of a checkpoint folder ischeckpoint-xx
convert_pytorch_to_roberta_original_pytorch_checkpoint.py
: we realize some probe like LAMA is most convenient to be run onfairseq
model, so we provide this converter script.