import asyncio import gradio as gr import pandas as pd import src.constants as constants from src.hub import glob, load_jsonlines_file def update_task_description_component(task): base_description = constants.TASK_DESCRIPTIONS.get(task, "") additional_info = "A higher score is a better score." description = f"{base_description}\n\n{additional_info}" if base_description else additional_info return gr.Textbox( description, label="Task Description", lines=6, visible=True, ) def update_subtasks_component(task, profile: gr.OAuthProfile | None): visible_login_btn = True if task == "leaderboard_gpqa" else False subtasks = None if task == "leaderboard_gpqa" and not profile else constants.SUBTASKS.get(task) return ( gr.LoginButton(size="sm", visible=visible_login_btn), gr.Radio( choices=subtasks, info="Evaluation subtasks to be loaded", value=None, ), ) def update_load_details_component(model_id, subtask): if model_id and subtask: return gr.Button("Load Details", interactive=True) else: return gr.Button("Load Details", interactive=False) def fetch_details_paths(model_id, subtask): model_name_sanitized = model_id.replace("/", "__") dataset_id = constants.DETAILS_DATASET_ID.format(model_name_sanitized=model_name_sanitized) filename = constants.DETAILS_FILENAME.format(subtask=subtask) path = f"{dataset_id}/**/{filename}" return glob(path) async def load_details_dataframe(model_id, subtask): if not model_id or not subtask: return paths = await asyncio.to_thread(fetch_details_paths, model_id, subtask) if not paths: return path = max(paths) data = await load_jsonlines_file(path) if not data: return df = pd.json_normalize(data) # Keep model_name: df["model_name"] = model_id return df.sort_values("doc_id").set_index("doc_id", drop=False).set_index("model_name", append=True) async def load_details(subtask, *model_ids_lists): dfs = await asyncio.gather( *[ load_details_dataframe(model_id, subtask) for model_ids in model_ids_lists if model_ids for model_id in model_ids ] ) dfs = [df for df in dfs if df is not None] if dfs: return pd.concat(dfs), None else: return None, None def display_details(df, sample_idx, show_only_differences): if df is None: return df = df.loc[df.index.levels[0][sample_idx]] df = df.T.rename_axis(columns=None) # Style # - Option: Show only differences any_difference = pd.Series(False, index=df.index) if show_only_differences: any_difference = df.ne(df.iloc[:, 0], axis=0).any(axis=1) return ( df.style.format(escape="html", na_rep="") # .hide(axis="index") # Hide non-different rows .hide([row for row in df.index if show_only_differences and not any_difference[row]]) # Fix overflow .set_table_styles( [ { "selector": "td", "props": [("overflow-wrap", "break-word"), ("max-width", "1px")], }, { "selector": ".col_heading", "props": [("width", f"{100 / len(df.columns)}%")], }, ] ) .to_html() ) def update_sample_idx_component(df): if df is None: return maximum = len(df) - 1 return gr.Number( label="Sample Index", info="Index of the sample to be displayed", value=0, minimum=0, maximum=maximum, visible=True, ) def clear_details(): # model_ids, details_dataframe, details_task, subtask, load_details_btn, sample_idx return ( gr.Dropdown(value=[]), None, None, None, gr.Button("Load Details", interactive=False), gr.Number(label="Sample Index", info="Index of the sample to be displayed", value=0, minimum=0, visible=False), ) def display_loading_message_for_details(): return "