eve-esa/mcqa-multiple-answers

Dataset Summary

EVE-mcqa-multiple-answers is a Multiple-Choice Question Answering (MCQA) dataset designed to evaluate the performance of language models in the domain of Earth Observation (EO). The dataset consists of questions related to EO concepts, technologies, and applications, each accompanied by multiple answer choices, with one or more correct answer.

Dataset Structure

Each example in the dataset contains an arbitrary number of possible choices and one or more correct answers. The structure of each example is as follows:

• Question: String representing the question related to Earth Observation.
• Answers: The list of the correct answer(s) indicated.
• Choices: A list of possible dictionary containing:
  • text: The text of the answer choice.
  • label: The label of the answer choice (e.g., "A", "B", "C", etc.).

Metrics

The metrics used to evaluate model performance on the EVE-mcqa dataset are:

• Exact Match (EM): Measures the percentage of perfect overlap between the predicted and the reference answer set.
• Intersection over Union (IoU): Measures the overlap between the predicted and reference answer sets.

Computing the Metrics

Here's how to compute both metrics in Python:

def exact_match(predicted_answers, reference_answers):
    """
    Compute Exact Match score.

    Args:
        predicted_answers: Set or list of predicted answer labels
        reference_answers: Set or list of reference answer labels

    Returns:
        1.0 if sets match exactly, 0.0 otherwise
    """
    pred_set = set(predicted_answers) if not isinstance(predicted_answers, set) else predicted_answers
    ref_set = set(reference_answers) if not isinstance(reference_answers, set) else reference_answers
    return 1.0 if pred_set == ref_set else 0.0

def intersection_over_union(predicted_answers, reference_answers):
    """
    Compute Intersection over Union (IoU) score.

    Args:
        predicted_answers: Set or list of predicted answer labels
        reference_answers: Set or list of reference answer labels

    Returns:
        IoU score between 0.0 and 1.0
    """
    pred_set = set(predicted_answers) if not isinstance(predicted_answers, set) else predicted_answers
    ref_set = set(reference_answers) if not isinstance(reference_answers, set) else reference_answers

    if len(pred_set) == 0 and len(ref_set) == 0:
        return 1.0

    intersection = len(pred_set & ref_set)
    union = len(pred_set | ref_set)

    return intersection / union if union > 0 else 0.0

# Example usage with the dataset
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("eve-esa/mcqa-multiple-answers", split="train")

# Example: Get predictions for a single example
example = dataset[0]
reference_answers = example["Answer"]  # Ground truth answer(s)

# Simulate model predictions (replace with your model's actual predictions)
predicted_answers = ["A", "C"]  # Your model's predicted answer labels

# Compute metrics
em_score = exact_match(predicted_answers, reference_answers)
iou_score = intersection_over_union(predicted_answers, reference_answers)

print(f"Question: {example['Question']}")
print(f"Reference answers: {reference_answers}")
print(f"Predicted answers: {predicted_answers}")
print(f"Exact Match: {em_score}")
print(f"IoU: {iou_score:.4f}")

# Compute average metrics across the entire dataset
total_em = 0
total_iou = 0

for example in dataset:
    # Replace this with your model's actual predictions
    predicted = your_model_predict(example["Question"], example["Choices"])
    reference = example["Answer"]

    total_em += exact_match(predicted, reference)
    total_iou += intersection_over_union(predicted, reference)

avg_em = total_em / len(dataset)
avg_iou = total_iou / len(dataset)

print(f"\nAverage Exact Match: {avg_em:.4f}")
print(f"Average IoU: {avg_iou:.4f}")

Citation

If you use this project in academic or research settings, please cite: