To fine-tune effectively, you need to format your data properly to provide clues to the model about where to start and stop generating text.
The indicator string is a symbol or sequence of symbols that you append to the end of your prompt to tell the model that you want it to start generating text after this string.
For example, if you want the model to categorize items as colors, you can use an indicator string like '->'. The prompts in your dataset would look like this:
You can use any string as an indicator string as long as it doesn't appear anywhere else in the dataset. We recommend using '\n###\n'.
The stop sequence is another special symbol or sequence of symbols that you use to tell the model that you want it to stop generating text after that point.
For example, if you want the model to generate one word as a completion, you can use a stop sequence such as "\n" (newline) or "." (period) to mark the end of the completion, like this:
'prompt' : 'banana ->', 'completion' : ' yellow \n'
'prompt' : 'lime ->', 'completion' : ' green \n'
'prompt' : 'tomato ->', 'completion' : ' red \n'
Calling the model
You should use the same symbols used in your dataset when calling the model. If you used the dataset above, you should use '\n' as a stop sequence. You should also append '->' to your prompts as an indicator string (e.g. prompt: 'lemon -> ')
It is important that you use consistent and unique symbols for the indicator string and the stop sequence, and that they don't appear anywhere else in your data. Otherwise, the model might get confused and generate unwanted or incorrect text.
We also recommend appending a single space character at the beginning of your outputs.
You can also use our command line tool to help format your dataset, after you have prepared it.