Exact prompt used to fine tune this model that will help when someone uses it for inference?
If someone were to use this model for generating responses based on a locally provided context (which seems to be very much compatible with the training data used for fine tuning with this model), it would be nice to know exactly how the text was formatted during fine tuning. Based on the pipeline.py script I found it's clear that it was close to this format:
INSTRUCTION_KEY = "### Instruction:"
RESPONSE_KEY = "### Response:"
END_KEY = "### End"
But the databricks-dolly-15k dataset used frequently adds a context, how exactly is that incorporated into the chunk of text when you fine tuned this model ?
You can see how the training input was formed with context or without here: https://github.com/databrickslabs/dolly/blob/master/training/trainer.py#L109
Thanks! I still had to dig into this file to verify exactly how the formatting worked in plaintext: https://github.com/databrickslabs/dolly/blob/master/training/consts.py
PROMPT_WITH_INPUT_FORMAT and how its constructed