Closed UestcJay closed 4 months ago
Hi @UestcJay, thanks for sharing this. The git log looks fine to me. After downloading all the part files, you'll need to run the following command in pretrain
folder and finetune
folder respectively, to combine the image packages into one:
cat images.tar.gz.part-* > images.tar.gz
This is because we split the images into multiple packages to make the uploading process more stable.
hi, I reproduce the traning with using the bunny dataset, the response of the model is neither an answer starting with yes
or no
when eval mme. Will your model be like this? how can i eval mme?
question:
Is this artwork created by gentile da fabriano? Please answer yes or no.
warning: Setting pad_token_id
to eos_token_id
:50256 for open-end generation.
response:
the artwork in question is not created by gentile da fabriano gentile da fabriano was an italian painter active
in the early renaissance, known for his work in florence the style of the painting, with its gold leaf background and the particular rendering of the figures, is more indicative of the work of artists from the late gothic period, such as fra angelico or giotto, who were active in the early 15th century the use of gold leaf and the specific iconography of the virgin mary and child are also more characteristic of the early renaissance, which followed the gothic period therefore, the correct answer to the question is no, this artwork is not created by gentile da fabriano
Please share more information.
And the warning shouldn't occur due to here. Please check your code version.
Okay, I probably forgot to add this line of code, I used the bunny data set for two stages of pretraining and full-parameter sft. The pretraining stage froze vit and llm, and the sft stage froze vit. The training strategies are the same. However, when I evaluated mme, there were problems like the example above. The model I trained would not answer starting with yes or no, the perception score is 1250, this should be abnormal, right? I don't know which step has the problem. Should all answer of your models start with yes or no?
Please use our code to train and evaluate the models.
"Should all answer of your models start with yes or no?" Yes.
Close the issue for now if there's no further discussions. Feel free to reopen it if there's any other questions.
Thanks for your great work! when I use modelscope python api to download training dataset, I failed:
when I use
git clone
directly, it shows:could you give me some advice? or can you upload to
huggingface
?