Jiawei Zhao

Transformer的Decoder部分

2

请问为什么不用transformer的decoder部分呢？

Hi weijun, congratulations! I'm really interested in your code and can't wait to try it out. So when are you going to release the code? Thank you! The paper was...

effect of identity matrix

1

Hi, what is the effect of the following operations? Why not just take the mean value of each class vector, like > out2=out2.mean(-1)  Thanks!

visualization

2

Great work! Could you share the visualization scripts for the following pic? Thanks! 

Support multiple rounds of video conversations?

4

Great work! As video conversations in the instruction dataset have only one round in this version, if I want to train and test multiple rounds of video conversions, what should...

DCNv3 ImportError: libc10.so: cannot open shared object file: No such file or directory

1

install DCNv3 with DCNv3-1.0+cu113torch1.12.1-cp38-cp38-linux_x86_64.whl my torch version: '1.12.1+cu113' import DCNv3 meets the following errors: ImportError: libc10.so: cannot open shared object file: No such file or directory  Looking for help!...

DPO训练报错KeyError: 'prompt_input_ids'

4

按自定义数据格式，训练DPO在Map时报错 File "ms-swift/swift/trainers/dpo_trainer.py", line 114, in tokenize_row if len(answer_tokens['prompt_input_ids']) + longer_response_length > self.max_length: KeyError: 'prompt_input_ids' 打印了下answer的key：dict_keys(['input_ids', 'attention_mask', 'prompt_inputs_embeds', 'prompt_attention_mask']) 训练代码： CUDA_VISIBLE_DEVICES=2 \ swift rlhf \ --rlhf_type dpo \ --model_type internvl2-4b...

sft是否支持指定模型的特定参数？

1

比如InternVL中的dynamic_image_size和max_dynamic_patch？