Enhancements and Fixes for Classification Functionality

[gokceuludogan]

I encountered several issues while using the library for classification tasks and implemented fixes alongside new features to enhance usability and flexibility.

Changes

• Updated Default Model Name
  • Replaced the default model name with the corresponding Hugging Face model name for better alignment and clarity.
• Fixed Missing Parameters in Trainer Initialization
  • Addressed missing parameters in the Evaluator initialization within TrainerForClassification.
  • Refactored the affected files:
    - turkish_lm_tuner/trainer.py
    - experiments/finetune.py
• Improved Dataset Integration
  • Updated DatasetProcessor to allow users to pass initialized datasets directly. This eliminates the need for library updates when integrating external datasets.
• Implemented New Metrics
  • Added support for calculating F1 metrics for multi label classification.

[gokceuludogan]

Minor changes made in new commits:

• Setting f1_multi_weighted as the default metric for multi-label classification
• Handling logits for multi-label classification
• Refactoring model initialization for multi-label classification

Significant changes:

• T5ForClassification implementation was not fully compatible with HuggingFace transformers, leading to issues while saving and loading weights. I have refactored the implementation to utilize T5Config so that it enables users to use HF conventions for initializing, loading and saving T5ForClassification. Trainer and evaluator classes have been refactored accordingly.

[zeynepyirmibes]

LGTM! Thank you for this correction