I can't figure out how to make fine tuning work
Hello everyone,
I'm seeking some expert guidance with a fine-tuning project and I'm hoping someone can help.
I have credits on this platform that I loaded specifically for fine-tuning a large language model. Unfortunately, I've had a lot of trouble getting the process to work correctly myself, and these credits are currently going unused.
My goal is to fine-tune the nothingiisreal/MN-12B-Celeste-V1.9 model from Hugging Face. I'm looking to use these datasets:
gboleda/wikicorpus_es
PleIAs/Spanish-PD-Books
ylacombe/google-argentinian-spanish
A critical constraint is that I need to be able to run the final fine-tuned model on my personal hardware: an NVIDIA RTX 3060 with 12GB of VRAM. This means the fine-tuning process will likely need to use techniques like quantization or LoRA to make the model small enough to fit.
Given these challenges and my specific requirements, I'm hoping someone might be able to offer a solution. Has anyone encountered similar issues or have suggestions on how I can move forward?
Could the platform's support team potentially perform this fine-tuning on my behalf using my existing credits, or are there any other solutions you would suggest?
Any advice or guidance would be greatly appreciated!
I'm seeking some expert guidance with a fine-tuning project and I'm hoping someone can help.
I have credits on this platform that I loaded specifically for fine-tuning a large language model. Unfortunately, I've had a lot of trouble getting the process to work correctly myself, and these credits are currently going unused.
My goal is to fine-tune the nothingiisreal/MN-12B-Celeste-V1.9 model from Hugging Face. I'm looking to use these datasets:
gboleda/wikicorpus_es
PleIAs/Spanish-PD-Books
ylacombe/google-argentinian-spanish
A critical constraint is that I need to be able to run the final fine-tuned model on my personal hardware: an NVIDIA RTX 3060 with 12GB of VRAM. This means the fine-tuning process will likely need to use techniques like quantization or LoRA to make the model small enough to fit.
Given these challenges and my specific requirements, I'm hoping someone might be able to offer a solution. Has anyone encountered similar issues or have suggestions on how I can move forward?
Could the platform's support team potentially perform this fine-tuning on my behalf using my existing credits, or are there any other solutions you would suggest?
Any advice or guidance would be greatly appreciated!
