Nerogar
[Feat] Extend metadata to save captions
Hello, can we please have more json metadata in the result LoRA's safetensors file? kohya-ss has lots of information like:
- trigger words,
- which model were used,
- how many images and it's bucketed resolutions,
- caption's words frequency,
- every parameter used for the trained model...
And maybe something else.
Agreed! It would be nice to have similar metadata such as what koyha_ss bakes in.
There already is an option to save the entire training configuration in the safetensors file. The only thing not included are the captions and trigger words, because those don't really exist within OneTrainer. The data loader is not set up to provide that kind of information
The data loader is not set up to provide that kind of information
actually that's what I miss - after if you train 10+ loras you will start to forget most trigger words after half a year and storing all that in separate files are not too handy. that's one of the reasons I stopped using OneTrainer.
There already is an option to save the entire training configuration in the safetensors file. The only thing not included are the captions and trigger words, because those don't really exist within OneTrainer. The data loader is not set up to provide that kind of information
I must have overlooked it even after looking at what I thought was all the documentation. Any hint on where?
EDIT: Duh! I found it :) Sorry!
The data loader is not set up to provide that kind of information
actually that's what I miss - after if you train 10+ loras you will start to forget most trigger words after half a year and storing all that in separate files are not too handy. that's one of the reasons I stopped using OneTrainer.
Agreed, the two main missing things for me are:
- Training tokens/text used (similar to what is in Koyha)
- Storing the sha256 hash of the base model used in training. Right now, it only saves the filename which while helpful, is not as definitive as a hash.