fix: missing arg

fix: typo

feat: restore state from checkpoint

Merge branch 'add-tokenizer-save' into feat-model

feat: model config not hardcoded

add tokenizer save to wandb:

fix: config used in preprocess

fix: custom reference

fix: typo

feat: allow loading a model checkpoint

feat: hardcode our full dataset

feat: bigger warnings

feat: use str mode for json

fix: import json

fix: output directory must exist

feat: update scriptst

feat: update model config + save optim

fix: model config

feat: use bart-large-cnn

fix: log metadata

fix: define function before it is used

fix: correct arg

feat: save model frequently

feat: split script for small and big runs

feat: update test script

feat: use bart large

fix: use correct key

fix: log correct metrics

feat: hardcode eval_steps

fix: eval_steps belongs to training_args

feat: eval_steps already exists in TrainingArguments

Merge branch 'main'

feat: set default x-axis

feat: log everything through wandb

Merge pull request #21 from borisdayma/feat-no_decay

feat: eval less often for faster training

Merge pull request #20 from borisdayma/eval-interval

feat: no decay option

feat: use common wandb shared folder

feat: change default for quick tests

feat: hardcoded datasets

Add eval_interval to evaluate and log every so often.

Shift tokens in numpy because the built in shift function stalls.

fix: should be converted to array

fix: labels array

fix: typo

fix: model config

fix: correct decoder_input_ids and labels

feat: don't log model by default

feat: fix typo