| 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) |
|
|
| |
| 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] |
|
|
| |
| 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 |
