[coati] How to get prompt_path and pretrain_dataset?

[gongel]

Hi, I want to reproduce the training process but have no two datasets. Do you have plans to open source datasets？
Thx.

ColossalAI/applications/Chat/examples/train_prompts.py

Lines 208 to 209 in 638a07a

	parser.add_argument('--prompt_path', type=str, default=None, help='path to the prompt dataset')
	parser.add_argument('--pretrain_dataset', type=str, default=None, help='path to the pretrained dataset')

[gongel]

And pretrain_dataset is sft dataset?

[binmakeswell]

@gongel We have released it. Thanks.
https://github.com/hpcaitech/ColossalAI/tree/main/applications/Chat#supervised-datasets-collection

[gongel]

OK. So pretrain_dataset is sft dataset and InstructionWild is sft dataset. For prompt_datasets, we can produce it from InstructionWild?

[guijuzhejiang]

I also have the same question,pretrain_dataset is sft dataset.But what dataset is prompt_path? Can the same be pretrain_dataset?

[guijuzhejiang]

Is it possible to extract a part of instructions in pretrain_dataset(sft dataset) to generate prompt_datasets?

[gongel]

Is it possible to extract a part of instructions in pretrain_dataset(sft dataset) to generate prompt_datasets?

I think so, too

[nctu6]

@gongel We have released it. Thanks. https://github.com/hpcaitech/ColossalAI/tree/main/applications/Chat#supervised-datasets-collection

In the init function of the class SupervisedDataset (sft_dataset.py),
there is one line of source code using the "output" key after loading the pretrain_dataset as below.
targets = [f"{example['output']}{tokenizer.eos_token}" for example in list_data_dict]

It will cause key errors when running train_prompts.py

Could the page describe the required column names in all data sets you used?

Thanks.

[XiaoLaoDi]

@gongel have you already solved this problem ?

[vincezengqiang]

Is it possible to extract a part of instructions in pretrain_dataset(sft dataset) to generate prompt_datasets?

I think so, too

hihi, 大佬好。想请教您一个问题，我们在训练sft以及prompt的时候，SupervisedDataset()会对instruction进行改写，但是我们在训练reward_model时没有进行改写，在inference.py中也没有进行改写。这样不同训练阶段的输入存在一定差别，不会对模型输出有影响么？

[SeekPoint]

@gongel We have released it. Thanks. https://github.com/hpcaitech/ColossalAI/tree/main/applications/Chat#supervised-datasets-collection

which file should be used