search

unslothai / unsloth

Finetune Llama 3.2, Mistral, Phi, Qwen 2.5 & Gemma LLMs 2-5x faster with 80% less memory
https://unsloth.ai
Apache License 2.0
18.42k stars 1.29k forks source link

Can i fine tune models that are not on unsloth/<model_name> list #1127

Closed miloskovacevic68 closed 1 month ago

miloskovacevic68 commented 1 month ago

Hello, recently I pretrained a QWEN2 1.5B model from scratch using my tokenizer and domain dataset. The model was downloaded from HF "Qwen/Qwen2-1.5B"
Can i use Unsloth to fine tune the model on my instruct dataset, or Unsloth can be used only with unsloth/? If the answer is yes could you suggest how to do that?

Thank you in advance, Milos

Sneakr commented 1 month ago

@miloskovacevic68 If the model architecture is supported by unsloth you can fine tune yes, it doesn't need to be unsloth/ repo. You can use the notebooks samples presented on the page and replace the model name.

If you want to train a base model, use an instruct template such as alpaca. If you train a fine tune or instruct tuned model, use the chat template from the original model as it was trained on.

miloskovacevic68 commented 1 month ago

Thank you very much.

@miloskovacevic68 If the model architecture is supported by unsloth you can fine tune yes, it doesn't need to be unsloth/ repo. You can use the notebooks samples presented on the page and replace the model name.

If you want to train a base model, use an instruct template such as alpaca. If you train a fine tune or instruct tuned model, use the chat template from the original model as it was trained on.

-- Reply to this email directly or view it on GitHub: https://github.com/unslothai/unsloth/issues/1127#issuecomment-2406032427 You are receiving this because you were mentioned.

Message ID: @.***>