Rurage

Updated repository information

**1. Automatic Ensemble Creation: Implement functionality for automatic creation of evaluation ensembles**

- Train your own RAG evaluation model (*i. e. Gradient Boosting Classifier*) based on the RuRAGE metrics
- Prepare data for the RuRAGE ensemble model training right from the RuRAGE reports
- Train your own RuRAGE ensemble model for any of the evaluation tasks: Correctness, Faithfulness, Relevance
- Save (only basic functionality is implemented) and use trained model on the inference

**2. Auto-adaptive thresholds: Implement functionality for automatic creation thresholds for features in ensemble**

- Automatic ensemble model optimization by selecting an optimal classification threshold(s)

**3. Multiclass Labels: Extend support to work with multiclass labels.**

- Train and inference RAG evaluation model to predict not only binary labels, but also multiclass labels for the Correctness/Faithfulness/Relevance estimation
- Automatic threshold selection supports multiclass labels

**4. Detailed usage examples**

- Notebook for the basic RuRAGE functionality (preparing metric reports for the Correctness, Faithfulness, Relevance tasks)
- Notebook for the RAG evaluation ensemble training and inferencing.

**5. Different bug fixes in the metric calculation**

1. Different types of RAG evaluation:
* __Correctness__ – how well does the model's response follow the context
* __Faithfulness__ – how well does the model follow the context
* __Relevance__ – how well does the model's answer relate to the question

2. Metrics to build evaluation

a) Deterministic Metrics:
* ROUGE
* BLEU
* Bigram overlap Precision
* Bigram overlap Recall
* Bigram overlap F1
* Unigram overlap Precision
* Unigram overlap Recall
* Unigram overlap F1

b) Model-based Metrics:
* NLI Scores using Transformer models
* Cosine Similarity using Transformer models