one of the things i changed today was the padding token in the middle of the text i changed it to the eos token i think this is because the setup complicated the generation of multiple outputs at once
it is performing worse with the change. no longer producing accurate completion on colab.
