data_loader.py

import pandas as pd
import numpy as np

RESULT_FILE = 'evaluation_results.xlsx'

metric_ud = {
    "Accuracy": 1,
    "Average Exact Match": 1,
    "Exact Match": 1,
    "F1 Score": 1,
    "AUC ROC": 1,
    "AUC PR": 1,
    "Precision": 1,
    "Recall": 1,
    "Equivalent": 1,
    "Bias": -1,
    "Toxicity": -1,
    "ROUGE-1": 1,
    "ROUGE-2": 1,
    "ROUGE-L": 1,
    "BLEU": 1,
    "SummaC": 1,
    "BERTScore": 1,
    "Coverage": 1,
    "Density": 1,
    "Compression": 1,
    "hLEPOR": 1,
    "Character Error Rate": -1,
    "Word Error Rate": -1,
    "Character Edit Distance": -1,
    "Word Edit Distance": -1,
    "Perplexity": -1,
    "Expected Calibration Error": -1,
    "acc@10": 1,
    "MRR@10 (Top 30)": 1,
    "NDCG@10 (Top 30)": 1,
    "MRR@10": 1,
    "NDCG@10": 1,
}

tasks = {
    "Information Retrieval": "informationretrieval",
    "Knowledge": "knowledge",
    "Language Modelling": "language-modelling",
    "Question Answering": "question-answering",
    "Reasoning": "reasoning",
    "Summarization": "summarization",
    "Text Classification": "text-classification",
    "Toxicity Detection": "toxicity-detection",
    "Translation": "translation",
    "Sentiment Analysis": "sentiment-analysis",
}

settings = {
    "Normal": "",
    "Few-shot Leanring": "fs",
    "Prompt Strategy 0": "pt0",
    "Prompt Strategy 1": "pt1",
    "Prompt Strategy 2": "pt2",
    "Chain-of-Thought": "cot",
    "Fairness": "fairness",
    "Robustness": "robustness",
}

task_w_settings = {
    "Information Retrieval": ["Normal", "Few-shot Leanring", "Robustness", "Fairness"],
    "Knowledge": ["Normal", "Few-shot Leanring", "Robustness"],
    "Language Modelling": ["Normal", "Few-shot Leanring", "Fairness"],
    "Question Answering": ["Prompt Strategy 0", "Prompt Strategy 1", "Prompt Strategy 2", "Robustness", "Fairness"],
    "Reasoning": ["Few-shot Leanring", "Chain-of-Thought"],
    "Summarization": ["Prompt Strategy 0", "Prompt Strategy 1", "Prompt Strategy 2", "Robustness"],
    "Text Classification": ["Normal", "Few-shot Leanring", "Robustness", "Fairness"],
    "Toxicity Detection": ["Normal", "Few-shot Leanring", "Robustness", "Fairness"],
    "Translation": ["Few-shot Leanring", "Robustness"],
    "Sentiment Analysis": ["Normal", "Few-shot Leanring", "Robustness", "Fairness"],
}

datasets = {
    "question-answering": {
        "xquad_xtreme": "xQUAD EXTREME",
        "mlqa": "MLQA",
    },
    "summarization": {
        "vietnews": "VietNews",
        "wikilingua": "WikiLingua",
    },
    "text-classification": {
        "vsmec": "VSMEC",
        "phoatis": "PhoATIS",
    },
    "toxicity-detection": {
        "victsd": "UIT-ViCTSD",
        "vihsd": "UIT-ViHSD",
    },
    "translation": {
        "phomt-envi": "PhoMT English-Vietnamese",
        "phomt-vien": "PhoMT Vietnamese-English",
        "opus100-envi": "OPUS-100 English-Vietnamese",
        "opus100-vien": "OPUS-100 Vietnamese-English",
    },
    "sentiment-analysis": {
        "vlsp": "VLSP 2016",
        "vsfc": "UIT-VSFC",
    },
    "informationretrieval": {
        "mmarco": "mMARCO",
        "mrobust": "mRobust",
    },
    "knowledge": {
        "zaloe2e": "ZaloE2E",
        "vimmrc": "ViMMRC",
    },
    "language-modelling": {
        "mlqa": "MLQA",
        "vsec": "VSEC",
    },
    "reasoning": {
        "srnatural-azr": "Synthetic Reasoning (Natural) - Azure",
        "srnatural-gcp": "Synthetic Reasoning (Natural) - Google Cloud",
        "srabstract-azr": "Synthetic Reasoning (Abstract Symbol)- Azure",
        "srabstract-gcp": "Synthetic Reasoning (Abstract Symbol)- Google Cloud",
        "math-azr": "MATH Level 1 - Azure",
        "math-gcp": "MATH Level 1 - Google Cloud",
    },
}


def load_data(file_name):
    """
    Load the data from the csv file
    """
    data = pd.read_excel(
        file_name,
        sheet_name=None,
        header=None
    )
    results = {}
    for task_name, task_id in tasks.items():
        for setting_name in task_w_settings[task_name]:
            setting_id = settings[setting_name]
            sheet_name = f"{task_id}-{setting_id}" if setting_id else task_id
            sheet_data = data[sheet_name]
            results_by_dataset = {}

            # Find the rows that contain the dataset ids
            # dataset_ids = datasets[task_id].keys()
            row_ids = []
            for i, row in sheet_data.iterrows():
                if "Models/" in row[0]:
                    row_ids.append(i)
            row_ids.append(len(sheet_data))

            # Get the data for each dataset
            for i in range(len(row_ids) - 1):
                dataset_id = sheet_data.iloc[row_ids[i]][0].split('/')[-1]
                dataset_name = datasets[task_id][dataset_id]

                dataset_data = sheet_data.iloc[row_ids[i] + 1: row_ids[i + 1]]
                dataset_data = dataset_data.fillna('')
                # dataset_data = dataset_data.dropna(axis=1, how='all')
                # dataset_data = dataset_data.dropna(axis=0, how='all')
                header = sheet_data.iloc[0]
                header[0] = "Models"

                # Create new pandas dataframe
                dataset_data = pd.DataFrame(
                    dataset_data.values, columns=header)
                results_by_dataset[dataset_name] = dataset_data

            results[f"{task_id}-{setting_id}"] = results_by_dataset

    return results


resutls = load_data(RESULT_FILE)