kcobbe
commented
2 years ago We don't provide code for training verifiers. The data in example_model_solutions.jsonl is simply a showcase of the results we achieve after running our own finetuning and verification pipelines. It is not training data.
If you want to train verifiers, you'll have to write that code yourself. As such, you're free to use any key names you like.
@kcobbe @vineetsk10 The blog post mentions about 'Verifiers' that improve the model performance. In the data , for instance there is this one which indicates that answer is not correct ==> "6b_verification": {"is_correct": false, etc...] which I believe are the verifiers..
when we create new datasets to fine tune the model, should we by any convention name a verifier key as "is_correct" ? or can it be anything arbitrary like "correct_answer: false" that GPT3 (api) would understand ?