Suppose a scenario:
I have a dataset that I want to finetune chatgpt3 on. I finetune the model davinci. I am charged with the training charged as per num token, epochs and davinci charges.
Now, I have a finetuned model with me.
How will I now be charged for the usage of this finetuned model? Will it be based on
- Number of requests?
- Number of tokens in completions generated?
- Number of tokens in prompts?
- Number of tokens in prompts + completions?
If yes, then what is the pricing like?
Total tokens input+output as far as I know, exact price depends on model, check out the pricing page
1 Like
UNTRAINED:
davinci untrained is 0.02 per 1000 tokens for usage
TRAINED:
davinci trained is 0.12 per 1000 tokens (6 times the price)
training davinci costs 0.03 per 1000 tokens
USING EMBEDDING INSTEAD:
Embedding is 0.0004 per 1000 tokens (to create the data - equivalent of training)
Usage is the cost of Davinci - but you have to include the text you find as a context.
Even if you fill the entire 4096 tokens, the Davinci request will be a total 0.08 for 4000 tokens (3/4 the cost of using 1000 tokens in a trained Davinci)
2 Likes
Thanks for your response.
Thanks for your response!