Using Hugging Face datasets for auto-regressive LM#
Overview#
This guide explains how to utilize Hugging Face datasets for training auto-regressive language models, offering two primary methods: offline and online:
Offline: Convert Hugging Face dataset to an HDF5 format, and then use a HDF5DataProcessor with a Cerebras Model Zoo model, such as GptHDF5DataProcessor for GPT.
Online Use a Hugging Face dataset directly as an dataloader with Cerebras Model Zoo to train and evaluate a language model (such as GPT).
Offline conversion of Hugging Face dataset to HDF5 format#
When converting a Hugging Face dataset to an HDF5 format, we recommend iterating over the dataset and writing samples to h5 files, which are later used by the HDF5IterableDataset class inside a HDF5DataProcessor.
To write samples to h5 files, you can call the script convert_dataset_to_HDF5.py. For more information, refer to Converting a PyTorch dataset to HDF5 format for a detailed explanation of this function.
The Cerebras Model Zoo has two examples that showcase the conversion of Hugging Face datasets. These examples contain two files. The first one, named Hugging Face_<example>.py defines a dataset using Hugging Face API. The second one, named HF_converter_example_<example>.py, that calls the convert_dataset_to_HDF5 function using the dataset defined by the first file.
Dataset |
Dataset definition |
Conversion to HDF5 |
Additional notes |
|---|---|---|---|
Based on this Hugging Face tutorial |
|||
Example of a Hugging Face dataset similar to Eli5 but larger |
Note
Cerebras GPT models expect the labels to be shifted in the dataloader rather than the model in contrast to Hugging Face. To address this difference, we have implemented a custom data collator function CSDataCollatorForLanguageModeling based on DataCollatorForLanguageModeling. Examples on using this function can be found in the examples HuggingFace_Eli5.py and HuggingFace_BookCorpus.py.
Once the samples are written into .h5 files, you will use a HDF5DataProcessor. For example, to train a GPT-style model, you will specify in the configuration yaml file:
train_input:
data_dir: <path to samples saved into h5 files>
data_processor: "GptHDF5DataProcessor"
...
eval_input:
data_dir: <path to samples saved into h5 files>
data_processor: "GptHDF5DataProcessor"
...
Online usage of Hugging Face dataset with Cerebras Model Zoo#
Alternatively, you can stream samples directly from your Hugging Face dataset without converting to HDF5. Since the preprocessing/tokenization can be CPU intensive and hurt the performance of the dataloader, we advise that users avoid this approach for large datasets.
The script file HuggingFaceDataProcessor.py provides the required tools to connect any Hugging Face dataset (Map-Style or Iterable) to CS GPT models.
As an example, the DataProcessor HuggingFaceDataProcessorEli5 class showcases the Hugging Face Eli5 (Map-Style) dataset directly streamed to GPT-2 model, as shown in this snippet
from modelzoo.data_preparation.huggingface.HuggingFace_Eli5 import (
HuggingFace_Eli5,
)
from modelzoo.data_preparation.huggingface.HuggingFaceDataProcessor import (
HuggingFaceDataProcessor,
)
from modelzoo.models.nlp.input_utils import num_tasks
class HuggingFaceDataProcessorEli5(HuggingFaceDataProcessor):
def __init__(self, params):
num_workers = params.get("num_workers", 0)
split = params["split"]
self.dataset, self.data_collator = HuggingFace_Eli5(
split=split, num_workers=num_workers
)
# The supper class will take care of sharding the dataset and creating the dataloader
super().__init__(params)
Notice that this code uses the HuggingFace_Eli5.py script to define the ELI5 dataset using HuggungFace API. Then it inherits from the HuggingFaceDataProcessor class.
Alternatively, Cerebras Model Zoo also includes the DataProcessor HuggingFaceIterableDataProcessorEli5 class as an example of using the same dataset in the Iterable format.
To use the DataProcessor HuggingFaceDataProcessorEli5 to train a GPT-style model, modify the configuration yaml file:
train_input:
data_processor: "HuggingFaceDataProcessorEli5"
split: "train"
batch_size: 128
shuffle: True
shuffle_seed: 1337
num_workers: 8
prefetch_factor: 10
persistent_workers: True # Important to avoid seeding at each epoch