Transformers documentation

How to fine-tune a model for common downstream tasks

Join the Hugging Face community

to get started

# How to fine-tune a model for common downstream tasks

This guide will show you how to fine-tune 🤗 Transformers models for common downstream tasks. You will use the 🤗 Datasets library to quickly load and preprocess the datasets, getting them ready for training with PyTorch and TensorFlow.

Before you begin, make sure you have the 🤗 Datasets library installed. For more detailed installation instructions, refer to the 🤗 Datasets installation page. All of the examples in this guide will use 🤗 Datasets to load and preprocess a dataset.

pip install datasets

Learn how to fine-tune a model for:

## Sequence classification with IMDb reviews

Sequence classification refers to the task of classifying sequences of text according to a given number of classes. In this example, learn how to fine-tune a model on the IMDb dataset to determine whether a review is positive or negative.

For a more in-depth example of how to fine-tune a model for text classification, take a look at the corresponding PyTorch notebook or TensorFlow notebook.

The 🤗 Datasets library makes it simple to load a dataset:

from datasets import load_dataset
imdb = load_dataset("imdb")

This loads a DatasetDict object which you can index into to view an example:

imdb["train"][0]
{'label': 1,
'text': 'Bromwell High is a cartoon comedy. It ran at the same time as some other programs about school life, such as "Teachers". My 35 years in the teaching profession lead me to believe that Bromwell High\'s satire is much closer to reality than is "Teachers". The scramble to survive financially, the insightful students who can see right through their pathetic teachers\' pomp, the pettiness of the whole situation, all remind me of the schools I knew and their students. When I saw the episode in which a student repeatedly tried to burn down the school, I immediately recalled ......... at .......... High. A classic line: INSPECTOR: I\'m here to sack one of your teachers. STUDENT: Welcome to Bromwell High. I expect that many adults of my age think that Bromwell High is far fetched. What a pity that it isn\'t!'
}

### Preprocess

The next step is to tokenize the text into a readable format by the model. It is important to load the same tokenizer a model was trained with to ensure appropriately tokenized words. Load the DistilBERT tokenizer with the AutoTokenizer because we will eventually train a classifier using a pretrained DistilBERT model:

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")

Now that you have instantiated a tokenizer, create a function that will tokenize the text. You should also truncate longer sequences in the text to be no longer than the model’s maximum input length:

def preprocess_function(examples):
return tokenizer(examples["text"], truncation=True)

Use 🤗 Datasets map function to apply the preprocessing function to the entire dataset. You can also set batched=True to apply the preprocessing function to multiple elements of the dataset at once for faster preprocessing:

tokenized_imdb = imdb.map(preprocess_function, batched=True)

Lastly, pad your text so they are a uniform length. While it is possible to pad your text in the tokenizer function by setting padding=True, it is more efficient to only pad the text to the length of the longest element in its batch. This is known as dynamic padding. You can do this with the DataCollatorWithPadding function:

from transformers import DataCollatorWithPadding
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)

### Fine-tune with the Trainer API

Now load your model with the AutoModelForSequenceClassification class along with the number of expected labels:

from transformers import AutoModelForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased", num_labels=2)

At this point, only three steps remain:

1. Define your training hyperparameters in TrainingArguments.
2. Pass the training arguments to a Trainer along with the model, dataset, tokenizer, and data collator.
3. Call Trainer.train() to fine-tune your model.
from transformers import TrainingArguments, Trainer

training_args = TrainingArguments(
output_dir='./results',
learning_rate=2e-5,
per_device_train_batch_size=16,
per_device_eval_batch_size=16,
num_train_epochs=5,
weight_decay=0.01,
)

trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_imdb["train"],
eval_dataset=tokenized_imdb["test"],
tokenizer=tokenizer,
data_collator=data_collator,
)

trainer.train()

### Fine-tune with TensorFlow

Fine-tuning with TensorFlow is just as easy, with only a few differences.

Start by batching the processed examples together with dynamic padding using the DataCollatorWithPadding function. Make sure you set return_tensors="tf" to return tf.Tensor outputs instead of PyTorch tensors!

from transformers import DataCollatorWithPadding
data_collator = DataCollatorWithPadding(tokenizer, return_tensors="tf")

Next, convert your datasets to the tf.data.Dataset format with to_tf_dataset. Specify inputs and labels in the columns argument:

tf_train_dataset = tokenized_imdb["train"].to_tf_dataset(
shuffle=True,
batch_size=16,
collate_fn=data_collator,
)

tf_validation_dataset = tokenized_imdb["train"].to_tf_dataset(
shuffle=False,
batch_size=16,
collate_fn=data_collator,
)

Set up an optimizer function, learning rate schedule, and some training hyperparameters:

from transformers import create_optimizer
import tensorflow as tf

batch_size = 16
num_epochs = 5
batches_per_epoch = len(tokenized_imdb["train"]) // batch_size
total_train_steps = int(batches_per_epoch * num_epochs)
optimizer, schedule = create_optimizer(
init_lr=2e-5,
num_warmup_steps=0,
num_train_steps=total_train_steps
)

Load your model with the TFAutoModelForSequenceClassification class along with the number of expected labels:

from transformers import TFAutoModelForSequenceClassification
model = TFAutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased", num_labels=2)

Compile the model:

import tensorflow as tf
model.compile(optimizer=optimizer)

Finally, fine-tune the model by calling model.fit:

model.fit(
tf_train_set,
validation_data=tf_validation_set,
epochs=num_train_epochs,
)

## Token classification with WNUT emerging entities

Token classification refers to the task of classifying individual tokens in a sentence. One of the most common token classification tasks is Named Entity Recognition (NER). NER attempts to find a label for each entity in a sentence, such as a person, location, or organization. In this example, learn how to fine-tune a model on the WNUT 17 dataset to detect new entities.

For a more in-depth example of how to fine-tune a model for token classification, take a look at the corresponding PyTorch notebook or TensorFlow notebook.

Load the WNUT 17 dataset from the 🤗 Datasets library:

from datasets import load_dataset
wnut = load_dataset("wnut_17")

A quick look at the dataset shows the labels associated with each word in the sentence:

wnut["train"][0]
{'id': '0',
'ner_tags': [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 7, 8, 8, 0, 7, 0, 0, 0, 0, 0, 0, 0, 0],
'tokens': ['@paulwalk', 'It', "'s", 'the', 'view', 'from', 'where', 'I', "'m", 'living', 'for', 'two', 'weeks', '.', 'Empire', 'State', 'Building', '=', 'ESB', '.', 'Pretty', 'bad', 'storm', 'here', 'last', 'evening', '.']
}

View the specific NER tags by:

label_list = wnut["train"].features[f"ner_tags"].feature.names
label_list
['O',
'B-corporation',
'I-corporation',
'B-creative-work',
'I-creative-work',
'B-group',
'I-group',
'B-location',
'I-location',
'B-person',
'I-person',
'B-product',
'I-product'
]

A letter prefixes each NER tag which can mean:

• B- indicates the beginning of an entity.
• I- indicates a token is contained inside the same entity (e.g., the State token is a part of an entity like Empire State Building).
• 0 indicates the token doesn’t correspond to any entity.

### Preprocess

Now you need to tokenize the text. Load the DistilBERT tokenizer with an AutoTokenizer:

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")

Since the input has already been split into words, set is_split_into_words=True to tokenize the words into subwords:

tokenized_input = tokenizer(example["tokens"], is_split_into_words=True)
tokens = tokenizer.convert_ids_to_tokens(tokenized_input["input_ids"])
tokens
['[CLS]', '@', 'paul', '##walk', 'it', "'", 's', 'the', 'view', 'from', 'where', 'i', "'", 'm', 'living', 'for', 'two', 'weeks', '.', 'empire', 'state', 'building', '=', 'es', '##b', '.', 'pretty', 'bad', 'storm', 'here', 'last', 'evening', '.', '[SEP]']

The addition of the special tokens [CLS] and [SEP] and subword tokenization creates a mismatch between the input and labels. Realign the labels and tokens by:

1. Mapping all tokens to their corresponding word with the word_ids method.
2. Assigning the label -100 to the special tokens [CLS] and “[SEP] so the PyTorch loss function ignores them.
3. Only labeling the first token of a given word. Assign -100 to the other subtokens from the same word.

Here is how you can create a function that will realign the labels and tokens:

def tokenize_and_align_labels(examples):
tokenized_inputs = tokenizer(examples["tokens"], truncation=True, is_split_into_words=True)

labels = []
for i, label in enumerate(examples[f"ner_tags"]):
word_ids = tokenized_inputs.word_ids(batch_index=i)  # Map tokens to their respective word.
previous_word_idx = None
label_ids = []
for word_idx in word_ids:                            # Set the special tokens to -100.
if word_idx is None:
label_ids.append(-100)
elif word_idx != previous_word_idx:              # Only label the first token of a given word.
label_ids.append(label[word_idx])

labels.append(label_ids)

tokenized_inputs["labels"] = labels
return tokenized_inputs

Now tokenize and align the labels over the entire dataset with 🤗 Datasets map function:

tokenized_wnut = wnut.map(tokenize_and_align_labels, batched=True)

Finally, pad your text and labels, so they are a uniform length:

from transformers import DataCollatorForTokenClassification
data_collator = DataCollatorForTokenClassification(tokenizer)

### Fine-tune with the Trainer API

Load your model with the AutoModelForTokenClassification class along with the number of expected labels:

from transformers import AutoModelForTokenClassification, TrainingArguments, Trainer
model = AutoModelForTokenClassification.from_pretrained("distilbert-base-uncased", num_labels=len(label_list))

Gather your training arguments in TrainingArguments:

training_args = TrainingArguments(
output_dir='./results',
evaluation_strategy="epoch",
learning_rate=2e-5,
per_device_train_batch_size=16,
per_device_eval_batch_size=16,
num_train_epochs=3,
weight_decay=0.01,
)

Collect your model, training arguments, dataset, data collator, and tokenizer in Trainer:

trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_wnut["train"],
eval_dataset=tokenized_wnut["test"],
data_collator=data_collator,
tokenizer=tokenizer,
)

trainer.train()

### Fine-tune with TensorFlow

from transformers import DataCollatorForTokenClassification
data_collator = DataCollatorForTokenClassification(tokenizer, return_tensors="tf")

Convert your datasets to the tf.data.Dataset format with to_tf_dataset:

tf_train_set = tokenized_wnut["train"].to_tf_dataset(
shuffle=True,
batch_size=16,
collate_fn=data_collator,
)

tf_validation_set = tokenized_wnut["validation"].to_tf_dataset(
shuffle=False,
batch_size=16,
collate_fn=data_collator,
)

Load the model with the TFAutoModelForTokenClassification class along with the number of expected labels:

from transformers import TFAutoModelForTokenClassification
model = TFAutoModelForTokenClassification.from_pretrained("distilbert-base-uncased", num_labels=len(label_list))

Set up an optimizer function, learning rate schedule, and some training hyperparameters:

from transformers import create_optimizer

batch_size = 16
num_train_epochs = 3
num_train_steps = (len(tokenized_datasets["train"]) // batch_size) * num_train_epochs
optimizer, lr_schedule = create_optimizer(
init_lr=2e-5,
num_train_steps=num_train_steps,
weight_decay_rate=0.01,
num_warmup_steps=0,
)

Compile the model:

import tensorflow as tf
model.compile(optimizer=optimizer)

Call model.fit to fine-tune your model:

model.fit(
tf_train_set,
validation_data=tf_validation_set,
epochs=num_train_epochs,
)

There are many types of question answering (QA) tasks. Extractive QA focuses on identifying the answer from the text given a question. In this example, learn how to fine-tune a model on the SQuAD dataset.

For a more in-depth example of how to fine-tune a model for question answering, take a look at the corresponding PyTorch notebook or TensorFlow notebook.

from datasets import load_dataset
squad = load_dataset("squad")

Take a look at an example from the dataset:

squad["train"][0]
'context': 'Architecturally, the school has a Catholic character. Atop the Main Building\'s gold dome is a golden statue of the Virgin Mary. Immediately in front of the Main Building and facing it, is a copper statue of Christ with arms upraised with the legend "Venite Ad Me Omnes". Next to the Main Building is the Basilica of the Sacred Heart. Immediately behind the basilica is the Grotto, a Marian place of prayer and reflection. It is a replica of the grotto at Lourdes, France where the Virgin Mary reputedly appeared to Saint Bernadette Soubirous in 1858. At the end of the main drive (and in a direct line that connects through 3 statues and the Gold Dome), is a simple, modern stone statue of Mary.',
'id': '5733be284776f41900661182',
'question': 'To whom did the Virgin Mary allegedly appear in 1858 in Lourdes France?',
'title': 'University_of_Notre_Dame'
}

### Preprocess

Load the DistilBERT tokenizer with an AutoTokenizer:

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")

There are a few things to be aware of when preprocessing text for question answering:

1. Some examples in a dataset may have a very long context that exceeds the maximum input length of the model. You can deal with this by truncating the context and set truncation="only_second".
2. Next, you need to map the start and end positions of the answer to the original context. Set return_offset_mapping=True to handle this.
3. With the mapping in hand, you can find the start and end tokens of the answer. Use the sequence_ids method to find which part of the offset corresponds to the question, and which part of the offset corresponds to the context.

Assemble everything in a preprocessing function as shown below:

def preprocess_function(examples):
questions = [q.strip() for q in examples["question"]]
inputs = tokenizer(
questions,
examples["context"],
max_length=384,
truncation="only_second",
return_offsets_mapping=True,
)

offset_mapping = inputs.pop("offset_mapping")
start_positions = []
end_positions = []

for i, offset in enumerate(offset_mapping):
sequence_ids = inputs.sequence_ids(i)

# Find the start and end of the context
idx = 0
while sequence_ids[idx] != 1:
idx += 1
context_start = idx
while sequence_ids[idx] == 1:
idx += 1
context_end = idx - 1

# If the answer is not fully inside the context, label it (0, 0)
if offset[context_start][0] > end_char or offset[context_end][1] < start_char:
start_positions.append(0)
end_positions.append(0)
else:
# Otherwise it's the start and end token positions
idx = context_start
while idx <= context_end and offset[idx][0] <= start_char:
idx += 1
start_positions.append(idx - 1)

idx = context_end
while idx >= context_start and offset[idx][1] >= end_char:
idx -= 1
end_positions.append(idx + 1)

inputs["start_positions"] = start_positions
inputs["end_positions"] = end_positions
return inputs

Apply the preprocessing function over the entire dataset with 🤗 Datasets map function:

tokenized_squad = squad.map(preprocess_function, batched=True, remove_columns=squad["train"].column_names)

Batch the processed examples together:

from transformers import default_data_collator
data_collator = default_data_collator

### Fine-tune with the Trainer API

from transformers import AutoModelForQuestionAnswering, TrainingArguments, Trainer
model = AutoModelForQuestionAnswering.from_pretrained("distilbert-base-uncased")

Gather your training arguments in TrainingArguments:

training_args = TrainingArguments(
output_dir='./results',
evaluation_strategy="epoch",
learning_rate=2e-5,
per_device_train_batch_size=16,
per_device_eval_batch_size=16,
num_train_epochs=3,
weight_decay=0.01,
)

Collect your model, training arguments, dataset, data collator, and tokenizer in Trainer:

trainer = Trainer(
model=model,
args=training_args,
data_collator=data_collator,
tokenizer=tokenizer,
)

trainer.train()

### Fine-tune with TensorFlow

Batch the processed examples together with a TensorFlow default data collator:

from transformers.data.data_collator import tf_default_collator
data_collator = tf_default_collator

Convert your datasets to the tf.data.Dataset format with the to_tf_dataset function:

tf_train_set = tokenized_squad["train"].to_tf_dataset(
dummy_labels=True,
shuffle=True,
batch_size=16,
collate_fn=data_collator,
)

dummy_labels=True,
shuffle=False,
batch_size=16,
collate_fn=data_collator,
)

Set up an optimizer function, learning rate schedule, and some training hyperparameters:

from transformers import create_optimizer

batch_size = 16
num_epochs = 2
total_train_steps = (len(tokenized_squad["train"]) // batch_size) * num_epochs
optimizer, schedule = create_optimizer(
init_lr=2e-5,
num_warmup_steps=0,
num_train_steps=total_train_steps,
)

from transformers import TFAutoModelForQuestionAnswering
model = TFAutoModelForQuestionAnswering("distilbert-base-uncased")

Compile the model:

import tensorflow as tf
model.compile(optimizer=optimizer)

Call model.fit to fine-tune the model:

model.fit(
tf_train_set,
validation_data=tf_validation_set,
epochs=num_train_epochs,
)`