Implement Med-Halt Tests for Robust Model Evaluation

[chakravarthik27]

Issue Description:
To enhance the robustness of model evaluation in Langtest, we propose implementing a suite of Med-Halt Tests. These tests aim to challenge the model's reliability and confidence by introducing variations in ground truth, adding distractors, and testing its ability to handle unexpected inputs.

Proposed Tests:

1. False Confidence Test (FCT):

   • Provide the model with deliberately false answers for evaluation.
   • Measure the model's confidence and its ability to avoid overconfidence in incorrect answers.

2. None of the Above Test (NOTA):

   • Introduce a "None of the Above" (NOTA) option to the question set.
   • Evaluate if the model can correctly identify when none of the provided answers are valid.

3. Fake Questions Test (FQT):

   • Select and modify random questions from the dataset.
   • Assess the model's ability to handle slightly altered or out-of-context questions effectively.

Benefits:

• Improves the model's reliability in handling diverse scenarios.
• Tests the model's ability to detect and reject incorrect or altered inputs.
• Enhances the overall robustness and trustworthiness of the evaluation process.

Acceptance Criteria:

• False Confidence Test (FCT) is implemented and validated.
• None of the Above Test (NOTA) functionality is added, with appropriate evaluation metrics.
• Fake Questions Test (FQT) is integrated and tested for effectiveness.
• Unit tests are added to verify the correctness of each Med-Halt test.
• Documentation is updated to describe the new test functionalities and their usage.