LLaMA-Factory
LLaMA-Factory copied to clipboard
Unify Efficient Fine-Tuning of 100+ LLMs
### Reminder - [X] I have read the README and searched the existing issues. ### Reproduction 1. https://github.com/abacusai/smaug 2. https://arxiv.org/abs/2402.13228 ### Expected behavior _No response_ ### System Info _No response_...
想请教下, 在gsm8k数据集上采取完全相同的设置,仅修改精度bf16,fp16 和 学习率 、epoch等 训练均已收敛,存在如下三种效果:(多次实验,很普遍的现象。训练和推理用的是同一个模版) ![f47b2665288d45292595c44da663255f](https://github.com/hiyouga/LLaMA-Factory/assets/43243253/2e2418fc-c263-4e7a-aa7b-8384442fa250) ![eece5e1823159f4c27a39bb0a7a93fd8](https://github.com/hiyouga/LLaMA-Factory/assets/43243253/e193bb26-b711-46af-8cfd-451fd22dd5a0) ![e9d98c1b6275031ce73e8521a0e78145](https://github.com/hiyouga/LLaMA-Factory/assets/43243253/15fd5329-bdc3-4501-a23c-1bbfe7245b86)
Dear Developers, I'm delighted to inform you that the documentation update for Python scripts has been successfully completed. The updated documentation provides clear explanations of function parameters, return types, and...
### Reminder - [X] I have read the README and searched the existing issues. ### Reproduction We observe that across models and datasets using Zero3 with bf16 yields much higher...
### Reminder - [X] I have read the README and searched the existing issues. ### Reproduction deepspeed --num_gpus 2 src/train_bash.py \ --deepspeed ds_config.json \ --stage sft \ --do_train \ --model_name_or_path...
https://huggingface.co/docs/trl/main/en/dpo_trainer#downsides-to-merging-qlora-before-dpo-approach-2 https://github.com/jondurbin/qlora/blob/main/qmerge.py
### Reminder - [X] I have read the README and searched the existing issues. ### Reproduction Multiturn datasets are supported in the original dpo example. Wouldn't it be appropriate for...
### Reminder - [X] I have read the README and searched the existing issues. ### Reproduction 看目前sft_packing的实现只是单纯将不同的单轮sft数据拼接到一起,然后分别计算target部分的loss def preprocess_packed_supervised_dataset( examples: Dict[str, List[Any]], tokenizer: "PreTrainedTokenizer", template: "Template", data_args: "DataArguments", ) ->...
### Reminder - [X] I have read the README and searched the existing issues. ### Reproduction I would like to know if there is an option to evaluate the model...