​
Installation

To train, make sure you install the training extra:

pip install model2vec[training]

​
Training a Classifier

Model2Vec supports training simple classifiers on top of static models. These models are extremely lightweight and can be trained on a CPU.

​
Initializing a Classifier

To initialize a classifier, you can use the Classifier class. This class allows you to create a lightweight model that can be trained on your data. We support both single- and multi-label classification, which work seamlessly based on the labels you provide.

from model2vec.distill import distill
from model2vec.train import StaticModelForClassification

# From a distilled model
distilled_model = distill("baai/bge-base-en-v1.5")
classifier = StaticModelForClassification.from_static_model(model=distilled_model)

# From a pre-trained model: potion is the default
classifier = StaticModelForClassification.from_pretrained(model_name="minishlab/potion-base-32m")

​
Load a Dataset

Next, load a dataset to train on.

from datasets import load_dataset

# Load the train and test splits of the dataset
train, test = load_dataset("setfit/subj", split=["train", "test"])

# Create X and y
X_train, y_train = train["text"], train["label"]
X_test, y_test   = test["text"], test["label"]

​
Training and Evaluating the Classifier

# Train the classifier
classifier = classifier.fit(X_train, y_train)

# Evaluate the classifier
results = classifier.evaluate(X_test, y_test)

​
Persistence

You can easily turn a trained classifier into a scikit-learn compatible pipeline and save it locally, or push it to the HuggingFace Hub:

pipeline = classifier.to_pipeline()

pipeline.save_pretrained(path)
pipeline.push_to_hub("my_org/my_classifier")

Then, you can load it back with:

from model2vec.inference import StaticModelPipeline

pipeline = StaticModelPipeline.from_pretrained("my_org/my_classifier")

Note that, since the converted pipeline is a scikit-learn pipeline, you don’t need Torch anymore for inference. This allows you to deploy your model in a lightweight environment without the need for heavy dependencies.