This is the dataset I obtained by applying instruction backtranslation (just the self-curation part, no self-augmentation).
The model used for curation is starcoder, fine-tuned on OpenAssistant-guanaco. Here is the command :
python3 -u -m torch.distributed.run main.py
--model_name_or_path=bigcode/starcoder
--dataset_name_or_path=ArmelR/oasst1_guanaco
--shuffle_buffer 100
--seq_length 2048
--max_steps 160
--batch_size 1… See the full description on the dataset page:
https://huggingface.co/datasets/ArmelR/test_instruction_backtranslation.
