Spaces:
Running
Running
Super neat ๐
#12
by pinkachu - opened
Your scripts are awesome!! Both of my Karcher merges tolerated ONE outlier so long as the cosine similarity and weight magnitude readings were good ๐ฆพ
Thanks, glad to hear you found them useful. Another thing you may want to check if merging Mistral Nemo models are the tokenizers (Mistral Tekken vs ChatML). For this I recommend eos_scanner.py
Update: A few more PMPF scripts were added for dataset management
https://huggingface.co/spaces/Naphula/model_tools/commit/a4c4002527ca01926e92fa60aa760b352875e5b1
https://huggingface.co/spaces/Naphula/model_tools/commit/6f2444f95c727d4ef53ff82926aaa99b05be669e
Ooo, new toys!! Thank You โจ