Trains a binary classifier on the given data stream.

Options

• :epochs - Number of training epochs. Defaults to 10.
• :batch_size - Batch size for training. Defaults to 32.
• :learning_rate - Initial learning rate. Defaults to 1.0e-3.
• :decay - Learning rate decay. Defaults to 1.0e-2.
• :labels - Mapping of labels. Can be a list or a map. Defaults to [0, 1].
• :validation_split - Fraction of data to use for validation. Defaults to 0.1.
• :patience - Number of epochs to wait for improvement before early stopping. Defaults to 5.
• :compiler - The Nx compiler to use. Defaults to EXLA.
• :model_version - The architecture version to use. Defaults to conservative_cnn.
• :tune - If true, performs automatic hyperparameter tuning for learning rate and dropout. Defaults to false.
• :dropout_rate - Dropout rate for the model (ignored if :tune is true). Defaults to 0.2.
• :false_positive_penalty - Penalty multiplier applied to validation false-positive rate when selecting checkpoints and tuned hyperparameters. Defaults to 0.5 for v7 and 0.0 for older models.
• :calibrate_threshold - If true, calibrates and persists a positive threshold on the validation split. Defaults to true for v7 and false for older models.
• :threshold_candidates - Thresholds to evaluate during calibration. Defaults to 0.5..0.9 in 0.05 increments.
• :vector_length - Fixed sequence length for tokenization. Defaults to 512.
• :tokenizer_data - Custom data stream to train the tokenizer. Defaults to the :text field of data_stream.