Open Wolfscowl opened 1 year ago
The same problem I had, have you solved it? What exactly caused it?
The same problem I had, have you solved it? What exactly caused it?
Unfortunately, I do not know. I could not solve the problem. If you get the problem solved in the near future, please let me know.
I find the same problem. In prompter.py
get_response() function, there is
return output.split(self.template["response_split"])[1].strip()
And the model repeats again and again, like this:
<unk>Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Instruction:
Recommend a movie
### Response:
I recommend the movie "The Godfather" because it is a great movie.
### Instruction:
Recommend a movie
### Response:
I recommend the movie "The Godfather" because it is a great movie.
### Instruction:
Recommend a movie
### Response:
I recommend the movie "The Godfather" because it is a great movie.
### Instruction:
Recommend a movie
### Response:
I recommend the movie "The Godfather" because it is a great movie.
### Instruction:
Recommend a movie
### Response:
I recommend the movie "The Godfather" because it is a great movie.
So, after split, we get:
I recommend the movie "The Godfather" because it is a great movie.
### Instruction:
Recommend a movie
I always get a weird output after the training when generating text (see screenshots)
The following tags are appended to the output 95% of the time: '### further instruction:' '### input:' '### output:' '### explanation:'
I use for training the default template (alpaca.json).
I can't figure out why the text generation generates such an output.
I use windows for training with
As checkpoint:
Training with following parameters (see screenshot).
I have already made countless different training sessions with different epochs, learningrates, datasets, ... Always the same confuse output attachment
I hope someone has an idea what could be the reason for this. Here are a few more examples: