Datasets documentation

Quick Start

Join the Hugging Face community

to get started

# Quick Start

The quick start is intended for developers who are ready to dive in to the code, and see an end-to-end example of how they can integrate 🤗 Datasets into their model training workflow. For beginners who are looking for a gentler introduction, we recommend you begin with the tutorials.

In the quick start, you will walkthrough all the steps to fine-tune BERT on a paraphrase classification task. Depending on the specific dataset you use, these steps may vary, but the general steps of how to load a dataset and process it are the same.

For more detailed information on loading and processing a dataset, take a look at Chapter 3 of the Hugging Face course! It covers additional important topics like dynamic padding, and fine-tuning with the Trainer API.

Get started by installing 🤗 Datasets:

pip install datasets

## Load the dataset and model

Begin by loading the Microsoft Research Paraphrase Corpus (MRPC) training dataset from the General Language Understanding Evaluation (GLUE) benchmark. MRPC is a corpus of human annotated sentence pairs used to train a model to determine whether sentence pairs are semantically equivalent.

>>> from datasets import load_dataset
>>> dataset = load_dataset('glue', 'mrpc', split='train')

Next, import the pre-trained BERT model and its tokenizer from the 🤗 Transformers library:

Pytorch
Hide Pytorch content
>>> from transformers import AutoModelForSequenceClassification, AutoTokenizer
>>> model = AutoModelForSequenceClassification.from_pretrained('bert-base-cased')
Some weights of the model checkpoint at bert-base-cased were not used when initializing BertForSequenceClassification: ['cls.predictions.bias', 'cls.predictions.transform.dense.weight', 'cls.predictions.transform.dense.bias', 'cls.predictions.decoder.weight', 'cls.seq_relationship.weight', 'cls.seq_relationship.bias', 'cls.predictions.transform.LayerNorm.weight', 'cls.predictions.transform.LayerNorm.bias']
- This IS expected if you are initializing BertForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPretraining model).
- This IS NOT expected if you are initializing BertForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
Some weights of BertForSequenceClassification were not initialized from the model checkpoint at bert-base-cased and are newly initialized: ['classifier.weight', 'classifier.bias']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
>>> tokenizer = AutoTokenizer.from_pretrained('bert-base-cased')
TensorFlow
Hide TensorFlow content
>>> from transformers import TFAutoModelForSequenceClassification, AutoTokenizer
>>> model = TFAutoModelForSequenceClassification.from_pretrained("bert-base-cased")
Some weights of the model checkpoint at bert-base-cased were not used when initializing TFBertForSequenceClassification: ['nsp___cls', 'mlm___cls']
- This IS expected if you are initializing TFBertForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPretraining model).
- This IS NOT expected if you are initializing TFBertForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
Some weights of TFBertForSequenceClassification were not initialized from the model checkpoint at bert-base-cased and are newly initialized: ['dropout_37', 'classifier']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
>>> tokenizer = AutoTokenizer.from_pretrained('bert-base-cased')

## Tokenize the dataset

The next step is to tokenize the text in order to build sequences of integers the model can understand. Encode the entire dataset with Dataset.map(), and truncate and pad the inputs to the maximum length of the model. This ensures the appropriate tensor batches are built.

>>> def encode(examples):

>>> dataset = dataset.map(encode, batched=True)
>>> dataset[0]
{'sentence1': 'Amrozi accused his brother , whom he called " the witness " , of deliberately distorting his evidence .',
'sentence2': 'Referring to him as only " the witness " , Amrozi accused his brother of deliberately distorting his evidence .',
'label': 1,
'idx': 0,
'input_ids': array([  101,  7277,  2180,  5303,  4806,  1117,  1711,   117,  2292, 1119,  1270,   107,  1103,  7737,   107,   117,  1104,  9938, 4267, 12223, 21811,  1117,  2554,   119,   102, 11336,  6732, 3384,  1106,  1140,  1112,  1178,   107,  1103,  7737,   107, 117,  7277,  2180,  5303,  4806,  1117,  1711,  1104,  9938, 4267, 12223, 21811,  1117,  2554,   119,   102]),
'token_type_ids': array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]),
'attention_mask': array([1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1])}

Notice how there are three new columns in the dataset: input_ids, token_type_ids, and attention_mask. These columns are the inputs to the model.

## Format the dataset

Depending on whether you are using PyTorch, TensorFlow, or JAX, you will need to format the dataset accordingly. There are three changes you need to make to the dataset:

Pytorch
Hide Pytorch content
1. Rename the label column to labels, the expected input name in BertForSequenceClassification:
>>> dataset = dataset.map(lambda examples: {'labels': examples['label']}, batched=True)
1. Retrieve the actual tensors from the Dataset object instead of using the current Python objects.
2. Filter the dataset to only return the model inputs: input_ids, token_type_ids, and attention_mask.

Dataset.set_format() completes the last two steps on-the-fly. After you set the format, wrap the dataset in torch.utils.data.DataLoader:

>>> import torch
>>> dataset.set_format(type='torch', columns=['input_ids', 'token_type_ids', 'attention_mask', 'labels'])
{'attention_mask': tensor([[1, 1, 1,  ..., 0, 0, 0],
[1, 1, 1,  ..., 0, 0, 0],
[1, 1, 1,  ..., 0, 0, 0],
...,
[1, 1, 1,  ..., 0, 0, 0],
[1, 1, 1,  ..., 0, 0, 0],
[1, 1, 1,  ..., 0, 0, 0]]),
'input_ids': tensor([[  101,  7277,  2180,  ...,     0,     0,     0],
[  101, 10684,  2599,  ...,     0,     0,     0],
[  101,  1220,  1125,  ...,     0,     0,     0],
...,
[  101, 16944,  1107,  ...,     0,     0,     0],
[  101,  1109, 11896,  ...,     0,     0,     0],
[  101,  1109,  4173,  ...,     0,     0,     0]]),
'label': tensor([1, 0, 1, 0, 1, 1, 0, 1]),
'token_type_ids': tensor([[0, 0, 0,  ..., 0, 0, 0],
[0, 0, 0,  ..., 0, 0, 0],
[0, 0, 0,  ..., 0, 0, 0],
...,
[0, 0, 0,  ..., 0, 0, 0],
[0, 0, 0,  ..., 0, 0, 0],
[0, 0, 0,  ..., 0, 0, 0]])}
TensorFlow
Hide TensorFlow content
1. Rename the label column to labels, the expected input name in TFBertForSequenceClassification:
>>> dataset = dataset.map(lambda examples: {'labels': examples['label']}, batched=True)
1. Retrieve the actual tensors from the Dataset object instead of using the current Python objects.
2. Filter the dataset to only return the model inputs: input_ids, token_type_ids, and attention_mask.

Dataset.set_format() completes the last two steps on-the-fly. After you set the format, wrap the dataset in tf.data.Dataset:

>>> import tensorflow as tf
>>> dataset.set_format(type='tensorflow', columns=['input_ids', 'token_type_ids', 'attention_mask', 'labels'])
>>> features = {x: dataset[x].to_tensor(default_value=0, shape=[None, tokenizer.model_max_length]) for x in ['input_ids', 'token_type_ids', 'attention_mask']}
>>> tfdataset = tf.data.Dataset.from_tensor_slices((features, dataset["labels"])).batch(32)
>>> next(iter(tfdataset))
({'input_ids': <tf.Tensor: shape=(32, 512), dtype=int32, numpy=
array([[  101,  7277,  2180, ...,     0,     0,     0],
[  101, 10684,  2599, ...,     0,     0,     0],
[  101,  1220,  1125, ...,     0,     0,     0],
...,
[  101,  1109,  2026, ...,     0,     0,     0],
[  101, 22263,  1107, ...,     0,     0,     0],
[  101,   142,  1813, ...,     0,     0,     0]], dtype=int32)>, 'token_type_ids': <tf.Tensor: shape=(32, 512), dtype=int32, numpy=
array([[0, 0, 0, ..., 0, 0, 0],
[0, 0, 0, ..., 0, 0, 0],
[0, 0, 0, ..., 0, 0, 0],
...,
[0, 0, 0, ..., 0, 0, 0],
[0, 0, 0, ..., 0, 0, 0],
[0, 0, 0, ..., 0, 0, 0]], dtype=int32)>, 'attention_mask': <tf.Tensor: shape=(32, 512), dtype=int32, numpy=
array([[1, 1, 1, ..., 0, 0, 0],
[1, 1, 1, ..., 0, 0, 0],
[1, 1, 1, ..., 0, 0, 0],
...,
[1, 1, 1, ..., 0, 0, 0],
[1, 1, 1, ..., 0, 0, 0],
[1, 1, 1, ..., 0, 0, 0]], dtype=int32)>}, <tf.Tensor: shape=(32,), dtype=int64, numpy=
array([1, 0, 1, 0, 1, 1, 0, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0, 1, 0, 1, 1, 1,
0, 1, 1, 1, 0, 0, 1, 1, 1, 0])>)

## Train the model

Pytorch
Hide Pytorch content

Lastly, create a simple training loop and start training:

>>> from tqdm import tqdm
>>> device = 'cuda' if torch.cuda.is_available() else 'cpu'
>>> model.train().to(device)
>>> for epoch in range(3):
...     for i, batch in enumerate(tqdm(dataloader)):
...         batch = {k: v.to(device) for k, v in batch.items()}
...         outputs = model(**batch)
...         loss = outputs[0]
...         loss.backward()
...         optimizer.step()
...         if i % 10 == 0:
...             print(f"loss: {loss}")
TensorFlow
Hide TensorFlow content

Lastly, compile the model and start training:

>>> loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(reduction=tf.keras.losses.Reduction.NONE, from_logits=True)
>>> model.fit(tfdataset, epochs=3)