src/populate.py

import json
import os

import pandas as pd
import numpy as np
from src.display.formatting import has_no_nan_values, make_clickable_model
from src.display.utils import AutoEvalColumn, EvalQueueColumn
from src.leaderboard.read_evals import get_raw_eval_results


CLOSED_MODELS = {
    "openai/GPT-4o": {"params": 72000, "license": "proprietary", "likes": 0, "model_type": "🔒 : closed"},
    "Claude-3.5-Sonnet": {"params": 72000, "license": "proprietary", "likes": 0, "model_type": "🔒 : closed"},
}


def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchmark_cols: list) -> pd.DataFrame:
    """Creates a dataframe from all the individual experiment results"""
    raw_data = get_raw_eval_results(results_path, requests_path)
    all_data_json = [v.to_dict() for v in raw_data]
    print(all_data_json)

    df = pd.DataFrame.from_records(all_data_json)
    # print(df.head(10))
    # for model_name, info in CLOSED_MODELS.items():
    #     if model_name not in df['Model'].values:
    #         df = pd.concat([df, pd.DataFrame([{
    #             "Model": model_name,
    #             "params": info["params"],
    #             "license": info["license"],
    #             "likes": info["likes"],
    #             "model_type": info["model_type"],
    #             "Precision (%)": 0,
    #             "Title search rate (%)": 0
    #         }])], ignore_index=True)

    def extract_first(value):
        if isinstance(value, (list, np.ndarray)):
            return value[0] if len(value) > 0 else 0
        elif isinstance(value, (int, float)):
            return value
        else:
            return 0

    df["(T1) Precision (%)"] = df["(T1) Precision (%)"].apply(extract_first)
    # 将数组转标量，空数组变为 0
    df["(T1) Precision (%)"] = df["(T1) Precision (%)"].apply(extract_first)
    df["(T1) Title Search Rate (%)"] = df["(T1) Title Search Rate (%)"].apply(extract_first)
    df["(T1) Overlap (%)"] = df["(T1) Overlap (%)"].apply(extract_first)
    df["(T1) Precision (First Author) (%)"] = df["(T1) Precision (First Author) (%)"].apply(extract_first)
    df["(T1) Overlap (First Author) (%)"] = df["(T1) Overlap (First Author) (%)"].apply(extract_first)
    # Task 2
    df["(T2) Similarity (%)"] = df["(T2) Similarity (%)"].apply(extract_first)
    df["(T2) Entail (TRUE %)"] = df["(T2) Entail (TRUE %)"].apply(extract_first)
    df["(T2) Entail (GPT-4o %)"] = df["(T2) Entail (GPT-4o %)"].apply(extract_first)
    df["(T2) ROUGE-1 (%)"] = df["(T2) ROUGE-1 (%)"].apply(extract_first)
    df["(T2) ROUGE-2 (%)"] = df["(T2) ROUGE-2 (%)"].apply(extract_first)
    df["(T2) ROUGE-L (%)"] = df["(T2) ROUGE-L (%)"].apply(extract_first)
    # Task 3
    df["(T3) Precision (%)"] = df["(T3) Precision (%)"].apply(extract_first)
    df["(T3) Title Search Rate (%)"] = df["(T3) Title Search Rate (%)"].apply(extract_first)
    df["(T3) Overlap (%)"] = df["(T3) Overlap (%)"].apply(extract_first)
    df["(T3) KPR (%)"] = df["(T3) KPR (%)"].apply(extract_first)
    df["(T3) ROUGE-1 (%)"] = df["(T3) ROUGE-1 (%)"].apply(extract_first)
    df["(T3) ROUGE-2 (%)"] = df["(T3) ROUGE-2 (%)"].apply(extract_first)
    df["(T3) ROUGE-L (%)"] = df["(T3) ROUGE-L (%)"].apply(extract_first)
    
    # 平均值列
    df["Average ⬆️"] = df[["(T1) Precision (%)",
                           "(T1) Title Search Rate (%)",
                           "(T1) Overlap (%)",
                           "(T1) Precision (First Author) (%)",
                           "(T1) Overlap (First Author) (%)",
                           "(T2) Similarity (%)",
                           "(T2) Entail (TRUE %)",
                           "(T2) Entail (GPT-4o %)",
                           "(T2) ROUGE-1 (%)",
                           "(T2) ROUGE-2 (%)",
                           "(T2) ROUGE-L (%)",
                           "(T3) Precision (%)",
                           "(T3) Title Search Rate (%)",
                           "(T3) Overlap (%)",
                           "(T3) KPR (%)",
                           "(T3) ROUGE-1 (%)",
                           "(T3) ROUGE-2 (%)",
                           "(T3) ROUGE-L (%)"]].mean(axis=1)

    # 排序
    df = df.sort_values(by=["Average ⬆️"], ascending=False)

    # 保留需要显示的列
    cols = [c for c in cols if c in df.columns]
    df = df[cols].round(2)

    # 如果 benchmark_cols 有列不在 df 中，忽略
    benchmark_cols = [c for c in benchmark_cols if c in df.columns]
    df[benchmark_cols] = df[benchmark_cols].fillna(0)

    # if benchmark_cols:
    #     df = df[has_no_nan_values(df, benchmark_cols)]

    # print(df.head(10))
    return df



def get_evaluation_queue_df(save_path: str, cols: list) -> list[pd.DataFrame]:
    """Creates the different dataframes for the evaluation queues requestes"""
    entries = [entry for entry in os.listdir(save_path) if not entry.startswith(".")]
    all_evals = []

    for entry in entries:
        if ".json" in entry:
            file_path = os.path.join(save_path, entry)
            with open(file_path) as fp:
                data = json.load(fp)

            data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])
            data[EvalQueueColumn.revision.name] = data.get("revision", "main")

            all_evals.append(data)
        elif ".md" not in entry:
            # this is a folder
            sub_entries = [e for e in os.listdir(f"{save_path}/{entry}") if os.path.isfile(e) and not e.startswith(".")]
            for sub_entry in sub_entries:
                file_path = os.path.join(save_path, entry, sub_entry)
                with open(file_path) as fp:
                    data = json.load(fp)

                data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])
                data[EvalQueueColumn.revision.name] = data.get("revision", "main")
                all_evals.append(data)

    pending_list = [e for e in all_evals if e["status"] in ["PENDING", "RERUN"]]
    running_list = [e for e in all_evals if e["status"] == "RUNNING"]
    finished_list = [e for e in all_evals if e["status"].startswith("FINISHED") or e["status"] == "PENDING_NEW_EVAL"]
    df_pending = pd.DataFrame.from_records(pending_list, columns=cols)
    df_running = pd.DataFrame.from_records(running_list, columns=cols)
    df_finished = pd.DataFrame.from_records(finished_list, columns=cols)
    return df_finished[cols], df_running[cols], df_pending[cols]