Training

LoRA Training Module

Handles training of domain-specific LoRA adapters.

class ual_adapter.training.trainer.TextDataset(texts: List[str], tokenizer: PreTrainedTokenizer, max_length: int = 512)

Bases: Dataset

Simple text dataset for training.

__init__(texts: List[str], tokenizer: PreTrainedTokenizer, max_length: int = 512)

Initialize text dataset.

Parameters:

• texts – List of training texts

• tokenizer – Tokenizer to use

• max_length – Maximum sequence length

class ual_adapter.training.trainer.LoRATrainer(base_model: PreTrainedModel, tokenizer: PreTrainedTokenizer, device: str = 'auto')

Bases: object

Trainer for LoRA adapters with domain specialization.

__init__(base_model: PreTrainedModel, tokenizer: PreTrainedTokenizer, device: str = 'auto')

Initialize LoRA trainer.

Parameters:

• base_model – Base model to fine-tune

• tokenizer – Tokenizer for the model

• device – Device to train on

train_adapter(adapter_name: str, training_texts: List[str], validation_texts: List[str] | None = None, rank: int = 16, alpha: float = 16.0, dropout: float = 0.1, target_modules: List[str] | None = None, learning_rate: float = 0.0001, num_epochs: int = 3, batch_size: int = 8, warmup_steps: int = 100, logging_steps: int = 10, save_steps: int = 500, output_dir: str | None = None, use_huggingface_trainer: bool = True, **kwargs) → Dict[str, Any]

Train a LoRA adapter on domain-specific data.

Parameters:

• adapter_name – Name for the adapter

• training_texts – List of training texts

• validation_texts – Optional validation texts

• rank – LoRA rank

• alpha – LoRA alpha parameter

• dropout – LoRA dropout

• target_modules – Target modules for LoRA

• learning_rate – Learning rate

• num_epochs – Number of training epochs

• batch_size – Batch size

• warmup_steps – Number of warmup steps

• logging_steps – Log every N steps

• save_steps – Save checkpoint every N steps

• output_dir – Directory to save checkpoints

• use_huggingface_trainer – Whether to use HF Trainer

• **kwargs – Additional training arguments

Returns:

Dictionary with training results

compute_perplexity(model: Module, texts: List[str], batch_size: int = 8) → float

Compute perplexity on a set of texts.

Parameters:

• model – Model to evaluate

• texts – Texts to compute perplexity on

• batch_size – Batch size for evaluation

Returns:

Perplexity value

LoRATrainer Class

class ual_adapter.training.trainer.LoRATrainer(base_model: PreTrainedModel, tokenizer: PreTrainedTokenizer, device: str = 'auto')

Trainer for LoRA adapters with domain specialization.

__init__(base_model: PreTrainedModel, tokenizer: PreTrainedTokenizer, device: str = 'auto')

Initialize LoRA trainer.

Parameters:

• base_model – Base model to fine-tune

• tokenizer – Tokenizer for the model

• device – Device to train on

train_adapter(adapter_name: str, training_texts: List[str], validation_texts: List[str] | None = None, rank: int = 16, alpha: float = 16.0, dropout: float = 0.1, target_modules: List[str] | None = None, learning_rate: float = 0.0001, num_epochs: int = 3, batch_size: int = 8, warmup_steps: int = 100, logging_steps: int = 10, save_steps: int = 500, output_dir: str | None = None, use_huggingface_trainer: bool = True, **kwargs) → Dict[str, Any]

Train a LoRA adapter on domain-specific data.

Parameters:

• adapter_name – Name for the adapter

• training_texts – List of training texts

• validation_texts – Optional validation texts

• rank – LoRA rank

• alpha – LoRA alpha parameter

• dropout – LoRA dropout

• target_modules – Target modules for LoRA

• learning_rate – Learning rate

• num_epochs – Number of training epochs

• batch_size – Batch size

• warmup_steps – Number of warmup steps

• logging_steps – Log every N steps

• save_steps – Save checkpoint every N steps

• output_dir – Directory to save checkpoints

• use_huggingface_trainer – Whether to use HF Trainer

• **kwargs – Additional training arguments

Returns:

Dictionary with training results

compute_perplexity(model: Module, texts: List[str], batch_size: int = 8) → float

Compute perplexity on a set of texts.

Parameters:

• model – Model to evaluate

• texts – Texts to compute perplexity on

• batch_size – Batch size for evaluation

Returns:

Perplexity value