feat: support frozen embeddings

[000alen]

Add embedding configuration options to TrainingConfig and implement loading of pretrained embeddings in Trainer class. Enhance model initialization to support freezing embeddings and ensure dimensional consistency when loading embeddings.

[Copilot]

Pull Request Overview

This PR adds support for loading and optionally freezing pretrained embeddings from Hugging Face models. The implementation allows models to initialize with pretrained embeddings (like GPT-2) and freeze them during training to preserve learned representations.

• Adds embedding configuration options to TrainingConfig (freeze_embeddings, load_pretrained_embeddings, pretrained_model_name)
• Implements pretrained embedding loading with dimensional consistency checks in the Trainer class
• Reorders initialization sequence to load data first, ensuring correct vocab_size before model creation

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

File	Description
train_launcher.py	Maps new embedding configuration parameters from config dict to TrainingConfig
secure_transformer/train.py	Implements embedding loading/freezing logic and reorders initialization sequence
secure_transformer/config.py	Adds new embedding-related configuration fields with defaults
configs/gpt2_frozen_config.yaml	Provides example configuration for using frozen GPT-2 embeddings

Comments suppressed due to low confidence (1)

secure_transformer/train.py:76

• [nitpick] The variable name 'trainable_params' could be more descriptive. Consider renaming to 'trainable_parameters' for consistency with the logging message that mentions 'trainable params'.

        trainable_params = [p for p in self.model.parameters() if p.requires_grad]