Closed theophilegervet closed 1 year ago
@theblackcat102 Would you have this augmented_latin_cyrillic_oasst_2023-03-27_v2.jsonl
file by any chance? @andreaskoepf said you might :)
@theophilegervet Our released RM oasst-rm-2-pythia-6.9b-epoch-1 which were also used for the RLHF tuning was not trained on the augmented data. The augmentation was an experiment to add several broken/very low quality replies, it is not really necessary. If you want to run first RM experiments you can use the file 2023-04-12_oasst_ready.trees.jsonl.gz
of OASST1.
While running reward model training with
and no changes to the config, I get the following error
Am I missing something? Should I download data first?