src/populate.py

import os
import pandas as pd

try:
    from src.display.utils import fields
except ImportError:
    try:
        from src.utils import fields
    except ImportError:
        try:
            from display.utils import fields
        except ImportError:
            from utils import fields

try:
    from src.leaderboard.read_evals import get_raw_eval_results
except ImportError:
    try:
        from src.read_evals import get_raw_eval_results
    except ImportError:
        try:
            from leaderboard.read_evals import get_raw_eval_results
        except ImportError:
            from read_evals import get_raw_eval_results


def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchmark_cols: list) -> pd.DataFrame:
    raw_data = get_raw_eval_results(results_path, requests_path)
    all_col_contents = fields()
    all_col_names = [c.name for c in all_col_contents]

    if not raw_data:
        return pd.DataFrame(columns=all_col_names)

    all_data_json = [v.to_dict() for v in raw_data]
    df = pd.DataFrame.from_records(all_data_json)

    # Ensure all expected columns exist and are strictly typed
    for col_content in all_col_contents:
        if col_content.name not in df.columns:
            df[col_content.name] = 0.0 if col_content.type == "number" else ""
        elif col_content.type == "number":
            df[col_content.name] = pd.to_numeric(df[col_content.name], errors="coerce").fillna(0.0)
        else:
            df[col_content.name] = df[col_content.name].fillna("").astype(str)

    df = df[all_col_names]

    # Sort ascending by risk/composite score (lower score = lower risk = rank higher)
    for col in all_col_names:
        if "composite" in col.lower() or "risk" in col.lower() or "score" in col.lower():
            df = df.sort_values(by=[col], ascending=True)
            break

    return df.reset_index(drop=True)


def get_top_3_eval_cards(results_path: str):
    raw_data = get_raw_eval_results(results_path, "")
    if not raw_data:
        return []
    sorted_res = sorted(raw_data, key=lambda x: x.composite_score)
    top_3 = []
    for r in sorted_res[:3]:
        top_3.append({
            "model_name": r.full_model,
            "org": r.org,
            "factuality": r.results.get("factuality", 0.0),
            "blind_fraction": r.results.get("blind_fraction", 0.0),
            "composite_score": r.composite_score,
            "run_id": f"REV-{r.revision}",
            "date": r.date,
        })
    return top_3