Script for Data Loading Testing

[fabianlim]

Best way to test the dataloading is to run the actual script (e.g., finetune) and then pull out the data samples that is created.

• doing this, we guarantee the samples are created in the same way.
• we can test this without torch.distributed

In this PR we provide

• utilities for extracting data samples from the training script
• a test_finetune function (and script) to extract out the data

See the below usage example

from scripts.test_dataloading import test_finetune

data = test_finetune(
    model_name_or_path='/path/to/model/',
    train_file='/path/to/train_file', 
    max_train_steps=2, # number of steps to run
)

then we can get the samples from data[0], data[1], and so on

# data[0]
{'input_ids': tensor([[100256, 100256, 100256,  ...,  74477,     13, 100257],
         [100256, 100256, 100256,  ...,   8082,     13, 100257],
         [100256, 100256, 100256,  ...,   5687,     13, 100257],
         ...,
         [100256, 100256, 100256,  ...,   3157,     13, 100257],
         [100256, 100256, 100256,  ...,    279,   8712, 100257],
         [100256, 100256, 100256,  ...,  67092,   1210, 100257]]),
 'labels': tensor([[  -100,   -100,   -100,  ...,  74477,     13, 100257],
         [  -100,   -100,   -100,  ...,   8082,     13, 100257],
         [  -100,   -100,   -100,  ...,   5687,     13, 100257],
         ...,
         [  -100,   -100,   -100,  ...,   3157,     13, 100257],
         [  -100,   -100,   -100,  ...,    279,   8712, 100257],
         [  -100,   -100,   -100,  ...,  67092,   1210, 100257]])}

[garrett361]

Looks good to me. Haven't use mocks before, they look super useful. We are going to use lora?

[fabianlim]

Looks good to me. Haven't use mocks before, they look super useful. We are going to use lora?

Not sure..