src/populate.py

import json
import os

import pandas as pd

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
import yaml
from sklearn.metrics import cohen_kappa_score
import numpy as np

TYPES = ["str", "number", "number", "number", "number", "number"]

def read_json(file_path: str) -> list[dict]:
    """
    Read a JSON/JSONL file and return its contents as a list of dictionaries.
    
    Parameters:
        file_path (str): The path to the JSON file.
        
    Returns:
        list[dict]: The contents of the JSON file as a list of dictionaries.
    """
    try:
        with open(file_path) as f:
            data = [json.loads(x) for x in f]
        return data
    except json.decoder.JSONDecodeError:
        with open(file_path) as f:
            data = json.load(f)
        return data

def pairwise_compare(
    evaluator1_dir: str,
    evaluator2_dir: str,
) -> tuple[float, float]:
    """
    Compare pairwise evaluators.

    Args:
        evaluator1_dir: The directory containing the responses from the first evaluator.
        evaluator2_dir: The directory containing the responses from the second evaluator.

    Returns:
        None
    """

    evaluator1_responses = read_json(evaluator1_dir)
    evaluator2_responses = read_json(evaluator2_dir)
    assert len(evaluator1_responses) == len(evaluator2_responses)
    evaluator1_winners = np.array(
        [response["winner"] for response in evaluator1_responses]
    )
    evaluator2_winners = np.array(
        [response["winner"] for response in evaluator2_responses]
    )
    acc = (evaluator1_winners == evaluator2_winners).mean().item()
    agreement = cohen_kappa_score(evaluator1_winners, evaluator2_winners)
    return acc, agreement


def pairwise_meta_eval(
    human_dir: str,
    model_dir: str,
    model_dir_swap: str
) -> dict[float]:
    """
    Evaluate a pairwise evaluator.

    Args:
        human_dir: The directory containing the human responses.
        model_dir: The directory containing the model responses.
        model_dir_swap: The directory containing the model responses with swapped inputs.

    Returns:
        dict[float]: The accuracy and agreement.
    """
    acc, agr = pairwise_compare(human_dir, model_dir)
    swap_acc, swap_agr = pairwise_compare(
        human_dir, model_dir_swap,
    )
    acc = (acc + swap_acc) / 2
    agr = (agr + swap_agr) / 2
    models_acc, models_agr = pairwise_compare(
        model_dir, model_dir_swap,
    )
    return acc, agr, models_acc, models_agr

def load_leaderboard() -> pd.DataFrame:
    """Loads the leaderboard from the file system"""
    with open("./data/models.yaml") as fp:
        models = yaml.safe_load(fp)

    predictions = {k: [] for k in ["Model", "Accuracy", "Agreement", "Self-Accuracy", "Self-Agreement"]}

    for model in models:
        fdir = model["fdir"]
        acc, agr, models_acc, models_agr = pairwise_meta_eval(
            f"./data/instrusum.json",
            f"./predictions/{fdir}.jsonl",
            f"./predictions/{fdir}_swap.jsonl"
        )
        predictions["Model"].append(model["name"])
        predictions["Accuracy"].append(acc)
        predictions["Agreement"].append(agr)
        predictions["Self-Accuracy"].append(models_acc)
        predictions["Self-Agreement"].append(models_agr)
    return pd.DataFrame(predictions).sort_values(by="Agreement", ascending=False).round(decimals=3)

            


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]

    df = pd.DataFrame.from_records(all_data_json)
    df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=False)
    df = df[cols].round(decimals=2)

    # filter out if any of the benchmarks have not been produced
    df = df[has_no_nan_values(df, benchmark_cols)]
    return raw_data, 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 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]