src/populate.py

import os

import pandas as pd

from src.display.utils import AutoEvalColumn


def get_leaderboard_df_crm(
    crm_results_path: str, accuracy_cols: list, ts_cols: list
) -> tuple[pd.DataFrame, pd.DataFrame]:
    """Creates a dataframe from all the individual experiment results"""
    use_case_flavor_mapping_df = pd.read_csv(os.path.join(crm_results_path, "hf_leaderboard_flavor_mapping.csv"))
    sf_finetuned_models = ["SF-TextBase 70B", "SF-TextBase 7B", "SF-TextSum"]
    # sf_finetuned_models = []

    leaderboard_accuracy_df = pd.read_csv(os.path.join(crm_results_path, "hf_leaderboard_accuracy.csv"))
    leaderboard_accuracy_df = leaderboard_accuracy_df[~leaderboard_accuracy_df["Model Name"].isin(sf_finetuned_models)]

    leaderboard_accuracy_df = leaderboard_accuracy_df.join(
        use_case_flavor_mapping_df[["Use Case Name", "Cost and Speed: Flavor"]].set_index("Use Case Name"),
        on="Use Case Name",
    )

    leaderboard_cost_df = pd.read_csv(os.path.join(crm_results_path, "hf_leaderboard_latency_cost.csv"))
    leaderboard_cost_df = leaderboard_cost_df[~leaderboard_cost_df["Model Name"].isin(sf_finetuned_models)]
    leaderboard_accuracy_df = leaderboard_accuracy_df.join(
        leaderboard_cost_df.set_index(["Model Name", "Cost and Speed: Flavor"]),
        on=["Model Name", "Cost and Speed: Flavor"],
    )

    leaderboard_ts_df = pd.read_csv(os.path.join(crm_results_path, "hf_leaderboard_ts.csv"))
    leaderboard_ts_crm_bias_df = pd.read_csv(os.path.join(crm_results_path, "hf_leaderboard_crm_bias.csv"))
    leaderboard_ts_df = leaderboard_ts_df[~leaderboard_ts_df["Model Name"].isin(sf_finetuned_models)]
    leaderboard_ts_df = leaderboard_ts_df.join(leaderboard_ts_crm_bias_df.set_index("Model Name"), on="Model Name")
    privacy_cols = leaderboard_ts_df[
        [
            "Privacy Zero-Shot Match Avoidance",
            "Privacy Zero-Shot Reveal Avoidance",
            "Privacy Five-Shot Match Avoidance",
            "Privacy Five-Shot Reveal Avoidance",
        ]
    ].apply(lambda x: x.str.rstrip("%").astype("float") / 100.0, axis=1)

    leaderboard_ts_df["Privacy"] = privacy_cols.mean(axis=1).transform(lambda x: "{:,.2%}".format(x))
    leaderboard_ts_df["Bias No CI"] = leaderboard_ts_df["CRM Fairness"].transform(lambda x: x.split(" ")[0])

    ts_lvl2_cols = leaderboard_ts_df[
        [
            "Safety",
            "Privacy",
            "Truthfulness",
            "Bias No CI",
        ]
    ].apply(lambda x: x.str.rstrip("%").astype("float") / 100.0, axis=1)
    leaderboard_ts_df["Trust & Safety"] = ts_lvl2_cols.mean(axis=1).transform(lambda x: "{:,.2%}".format(x))

    leaderboard_accuracy_df = leaderboard_accuracy_df.join(
        leaderboard_ts_df[ts_cols].set_index(["Model Name"]),
        on=["Model Name"],
    )

    leaderboard_accuracy_df = leaderboard_accuracy_df.sort_values(
        by=[AutoEvalColumn.use_case_name.name, AutoEvalColumn.accuracy_metric_average.name], ascending=[True, False]
    )
    leaderboard_accuracy_df = leaderboard_accuracy_df[accuracy_cols].round(decimals=2)
    return leaderboard_accuracy_df