Marcelo Díaz
Hello, I have the same doubt, did you figured it out?
Hello there. I am also interested in using with VLLM a 8/4 bits model trained with Unsloth. Currently, it works fine with 16 bits but requires too much VRAM. Is...
Hi, I'm getting the same issue. Any solution?
I am facing the same issue using train_on_responses_only with Qwen 2.5 7B, and the solution is using the DataCollatorForSeq2Seq as the data_collator as follows: ```py from trl import SFTTrainer from...
For fine-tuning models using `train_on_responses_only`, I achieved faster training results by enabling `packing=True`. I assume this is because the model processes fewer padding tokens when putting training samples together to...
> How does > > ``` > instruction_part = "user\n\n", > response_part = "assistant\n\n", > ``` > > work when potentially multiple instruction headers could be presented, eg a response...