We finetuned the data using jsonl with system/user/assistant, and the training loss is approaching 0.
Yet, when I selected an example from training set and pair with the same system message, changing temperature to 0 and TopP to 0.5, I am still unable to reproduce the training set
Welcome to the Forum!
Can you share details about your fine-tuning use case, i.e. what did you try to achieve with the fine-tuning?
1 Like
I am using the system/user/assistant format, where each time the user would ask a question and the assistant would give an answer,usually the question is 10-20 token and answer in the length 50 tokens
Consider splitting a portion of your fine-tuning dataset to use as a validation set during the fine-tuning process.
Comparing the loss rate with the data not used in fine-tuning can help verify reproducibility.