Hey, I want to finetune davinci. So when giving him a scientific plant name it will give me the flower color in this format → {red}.
I tried following some tutorials but without success
You don’t need to fine-tune. You can just prompt. The input to a base model must be phrased as a leading question, a “completion”.
Also, lower the API temperature so it doesn’t pick less-likely outputs. Also tell it how to handle inapplicable inputs.
1 Like
But from my understanding finetuning is less cheaper if I provide a long prompt each time it will be more expensive no?
Fine-tuning is not the tool for the task. If anything, you would use embeddings to store the relationship between the scientific name of a flower and its color if it doesn’t already answer correctly in the public version of ChatGPT.
1 Like
Fine-tunes have a much higher training and usage cost. Davinci finetuned is 6x more expensive than base, and 60x more expensive than gpt-3.5-turbo chat. And you’d be paying a ton to experiment on tuning to barely affect its behavior.
My prompt plus answer was 57 tokens. Dump 500 tokens of storytelling instructions onto chat model to get even more predicable output, at lower price.
You can also “invest” in prompt language to make it process a hundred at a time.
2 Likes
Thanks I misread the pricing page of openai which lead me to think that it’s cheaper.