Datasets:
how to transfer this format into SFT format？
#26
by
Barry1915
- opened
For instance, {instruction、 input、output、history}
I got the same question. How does the dataset fine-tune the Pythia model in the v1 paper specifically? Should we sample some threads and turn them into multi-turn conversation forms?
Please refer to the code in the Open-Assistant repository: https://github.com/LAION-AI/Open-Assistant/tree/main/model/model_training (trainer_sft.py)
The specific formatting used can be found in the dialogue_collator.py and formatting.py.