I’m currently fine-tuning a model with my SMS conversation data and have encountered an issue. Although my JSON file contains 167 lines, the fine-tuning process appears to exceed this line count. Could there be a step I’m overlooking that would cause the fine-tuning to run beyond the provided number of conversation lines? Any guidance would be greatly appreciated.
Here is the formula:
Steps = Epochs * Lines
You would expect 498 total steps
Cheers Curt. Is there an easy way I can get an estimate cost in tokens/money to train a model?
From my experience, it usually predicts the price beforehand. So very EASY
But rough price estimate is, if you need to estimate it yourself (for English):
Tokens = 0.75 * CharactersPerLine * NumberOfLines * NumberOfEpochs
Then on the pricing page, look at a chart similar to this:
Then compute your token units:
TokenUnits = Tokens/1000
Then depending on your model, in your case, your estimate for
gpt-3.5-turbo is max $0.0060 per 1k, or per TokenUnits.
So in your case, the max you would pay is:
MaxDollarsForTraining = 0.0060 * TokenUnits
Much appreciated Curt . Thanks again