Fine tuning dataset
#1
by
gyuwon12
- opened
Hello! Can you tell me what dataset you used for fine tuning? And I would appreciate it if I could know additional hyperparameters or config.
We used the self instruct methodology to generate a variety of instructions and then used chatgpt for a long, detailed answer style to elicit ORCA-style responses.
For multi-turn conversations, we further trained it by creating dialogs starting with the generated instructions mentioned above.
We haven't written to the model card yet, but we've used the tokens unused0 and unused1 as the user token and response token, respectively.