import gradio as gr import pandas as pd from apscheduler.schedulers.background import BackgroundScheduler from huggingface_hub import snapshot_download import src.envs as envs from src.envs import API, EVAL_REQUESTS_PATH, EVAL_RESULTS_PATH, QUEUE_REPO, REPO_ID, RESULTS_REPO import socket import src.display.about as about from src.display.css_html_js import custom_css import src.display.utils as utils import src.populate as populate from src.populate import get_evaluation_queue_df, get_leaderboard_df import src.submission.submit as submit import os import spacy_transformers TOKEN = os.environ.get("H4_TOKEN", None) print("TOKEN", TOKEN) import src.backend.run_eval_suite as run_eval_suite def ui_snapshot_download(repo_id, local_dir, repo_type, tqdm_class, etag_timeout): try: print(local_dir) snapshot_download(repo_id=repo_id, local_dir=local_dir, repo_type=repo_type, tqdm_class=tqdm_class, etag_timeout=etag_timeout) except Exception as e: restart_space() def restart_space(): envs.API.restart_space(repo_id=envs.REPO_ID, token=TOKEN) def init_space(): #dataset_df = get_dataset_summary_table(file_path='blog/Hallucination-Leaderboard-Summary.csv') if socket.gethostname() not in {'neuromancer'}: # sync model_type with open-llm-leaderboard ui_snapshot_download(repo_id=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30) ui_snapshot_download(repo_id=RESULTS_REPO, local_dir=EVAL_RESULTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30) raw_data, original_df = get_leaderboard_df(EVAL_RESULTS_PATH, EVAL_REQUESTS_PATH, utils.COLS, utils.BENCHMARK_COLS) finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df = get_evaluation_queue_df(EVAL_REQUESTS_PATH, utils.EVAL_COLS) return original_df, finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df original_df, finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df = init_space() def process_pending_evals(): if len(pending_eval_queue_df) == 0: print("No pending evaluations found.") return for _, eval_request in pending_eval_queue_df.iterrows(): # 根据具体的列名和数据格式调整 `eval_request` 的内容 print(f"Evaluating model: {eval_request['model']}") # 调用评估函数 run_eval_suite.run_evaluation( eval_request=eval_request, local_dir=envs.EVAL_RESULTS_PATH_BACKEND, results_repo=envs.RESULTS_REPO, batch_size=1, device=envs.DEVICE, no_cache=True, need_check=True, # 根据需要设定是否需要检查 write_results=True # 根据需要设定是否写入结果 ) print(f"Finished evaluation for model: {eval_request['model']}") # 在初始化完成后调用 original_df, finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df = init_space() process_pending_evals() # try: # print(envs.EVAL_REQUESTS_PATH) # snapshot_download( # repo_id=envs.QUEUE_REPO, local_dir=envs.EVAL_REQUESTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30 # ) # except Exception: # restart_space() # try: # print(envs.EVAL_RESULTS_PATH) # snapshot_download( # repo_id=envs.RESULTS_REPO, local_dir=envs.EVAL_RESULTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30 # ) # except Exception: # restart_space() # raw_data, original_df = populate.get_leaderboard_df(envs.RESULTS_REPO, envs.QUEUE_REPO, utils.COLS, utils.BENCHMARK_COLS) leaderboard_df = original_df.copy() ( finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df, ) = populate.get_evaluation_queue_df(envs.EVAL_REQUESTS_PATH, utils.EVAL_COLS) # Searching and filtering def update_table( hidden_df: pd.DataFrame, columns: list, type_query: list, precision_query: str, size_query: list, show_deleted: bool, query: str, ): filtered_df = filter_models(hidden_df, type_query, size_query, precision_query, show_deleted) filtered_df = filter_queries(query, filtered_df) df = select_columns(filtered_df, columns) return df def search_table(df: pd.DataFrame, query: str) -> pd.DataFrame: return df[(df[utils.AutoEvalColumn.dummy.name].str.contains(query, case=False))] def select_columns(df: pd.DataFrame, columns: list) -> pd.DataFrame: always_here_cols = [ utils.AutoEvalColumn.model_type_symbol.name, utils.AutoEvalColumn.model.name, ] # We use COLS to maintain sorting filtered_df = df[ always_here_cols + [c for c in utils.COLS if c in df.columns and c in columns] + [utils.AutoEvalColumn.dummy.name] ] return filtered_df def filter_queries(query: str, filtered_df: pd.DataFrame) -> pd.DataFrame: final_df = [] if query != "": queries = [q.strip() for q in query.split(";")] for _q in queries: _q = _q.strip() if _q != "": temp_filtered_df = search_table(filtered_df, _q) if len(temp_filtered_df) > 0: final_df.append(temp_filtered_df) if len(final_df) > 0: filtered_df = pd.concat(final_df) filtered_df = filtered_df.drop_duplicates( subset=[utils.AutoEvalColumn.model.name, utils.AutoEvalColumn.precision.name, utils.AutoEvalColumn.revision.name] ) return filtered_df def filter_models( df: pd.DataFrame, type_query: list, size_query: list, precision_query: list, show_deleted: bool ) -> pd.DataFrame: # Show all models # if show_deleted: # filtered_df = df # else: # Show only still on the hub models # filtered_df = df[df[utils.AutoEvalColumn.still_on_hub.name]] filtered_df = df type_emoji = [t[0] for t in type_query] filtered_df = filtered_df.loc[df[utils.AutoEvalColumn.model_type_symbol.name].isin(type_emoji)] filtered_df = filtered_df.loc[df[utils.AutoEvalColumn.precision.name].isin(precision_query + ["None"])] numeric_interval = pd.IntervalIndex(sorted([utils.NUMERIC_INTERVALS[s] for s in size_query])) params_column = pd.to_numeric(df[utils.AutoEvalColumn.params.name], errors="coerce") mask = params_column.apply(lambda x: any(numeric_interval.contains(x))) filtered_df = filtered_df.loc[mask] return filtered_df demo = gr.Blocks(css=custom_css) with demo: gr.HTML(about.TITLE) gr.Markdown(about.INTRODUCTION_TEXT, elem_classes="markdown-text") with gr.Tabs(elem_classes="tab-buttons") as tabs: with gr.TabItem("🏅 LLM Benchmark", elem_id="llm-benchmark-tab-table", id=0): with gr.Row(): with gr.Column(): with gr.Row(): search_bar = gr.Textbox( placeholder=" 🔍 Search for your model (separate multiple queries with `;`) and press ENTER...", show_label=False, elem_id="search-bar", ) with gr.Row(): shown_columns = gr.CheckboxGroup( choices=[ c.name for c in utils.fields(utils.AutoEvalColumn) if not c.hidden and not c.never_hidden and not c.dummy ], value=[ c.name for c in utils.fields(utils.AutoEvalColumn) if c.displayed_by_default and not c.hidden and not c.never_hidden ], label="Select columns to show", elem_id="column-select", interactive=True, ) with gr.Row(): deleted_models_visibility = gr.Checkbox( value=False, label="Show gated/private/deleted models", interactive=True ) with gr.Column(min_width=320): #with gr.Box(elem_id="box-filter"): filter_columns_type = gr.CheckboxGroup( label="Model types", choices=[t.to_str() for t in utils.ModelType], value=[t.to_str() for t in utils.ModelType], interactive=True, elem_id="filter-columns-type", ) filter_columns_precision = gr.CheckboxGroup( label="Precision", choices=[i.value.name for i in utils.Precision], value=[i.value.name for i in utils.Precision], interactive=True, elem_id="filter-columns-precision", ) filter_columns_size = gr.CheckboxGroup( label="Model sizes (in billions of parameters)", choices=list(utils.NUMERIC_INTERVALS.keys()), value=list(utils.NUMERIC_INTERVALS.keys()), interactive=True, elem_id="filter-columns-size", ) leaderboard_table = gr.components.Dataframe( value=leaderboard_df[ [c.name for c in utils.fields(utils.AutoEvalColumn) if c.never_hidden] + shown_columns.value + [utils.AutoEvalColumn.dummy.name] ], headers=[c.name for c in utils.fields(utils.AutoEvalColumn) if c.never_hidden] + shown_columns.value, datatype=utils.TYPES, elem_id="leaderboard-table", interactive=False, visible=True, column_widths=["2%", "33%"] ) # Dummy leaderboard for handling the case when the user uses backspace key hidden_leaderboard_table_for_search = gr.components.Dataframe( value=original_df[utils.COLS], headers=utils.COLS, datatype=utils.TYPES, visible=False, ) search_bar.submit( update_table, [ hidden_leaderboard_table_for_search, shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, deleted_models_visibility, search_bar, ], leaderboard_table, ) for selector in [shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, deleted_models_visibility]: selector.change( update_table, [ hidden_leaderboard_table_for_search, shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, deleted_models_visibility, search_bar, ], leaderboard_table, queue=True, ) with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=2): gr.Markdown(about.LLM_BENCHMARKS_TEXT, elem_classes="markdown-text") with gr.TabItem("🚀 Submit here! ", elem_id="llm-benchmark-tab-table", id=3): with gr.Column(): with gr.Row(): gr.Markdown(about.EVALUATION_QUEUE_TEXT, elem_classes="markdown-text") with gr.Column(): with gr.Accordion( f"✅ Finished Evaluations ({len(finished_eval_queue_df)})", open=False, ): with gr.Row(): finished_eval_table = gr.components.Dataframe( value=finished_eval_queue_df, headers=utils.EVAL_COLS, datatype=utils.EVAL_TYPES, row_count=5, ) with gr.Accordion( f"🔄 Running Evaluation Queue ({len(running_eval_queue_df)})", open=False, ): with gr.Row(): running_eval_table = gr.components.Dataframe( value=running_eval_queue_df, headers=utils.EVAL_COLS, datatype=utils.EVAL_TYPES, row_count=5, ) with gr.Accordion( f"⏳ Pending Evaluation Queue ({len(pending_eval_queue_df)})", open=False, ): with gr.Row(): pending_eval_table = gr.components.Dataframe( value=pending_eval_queue_df, headers=utils.EVAL_COLS, datatype=utils.EVAL_TYPES, row_count=5, ) with gr.Row(): gr.Markdown("# ✉️✨ Submit your model here!", elem_classes="markdown-text") with gr.Row(): with gr.Column(): model_name_textbox = gr.Textbox(label="Model name") revision_name_textbox = gr.Textbox(label="Revision commit", placeholder="main") model_type = gr.Dropdown( choices=[t.to_str(" : ") for t in utils.ModelType if t != utils.ModelType.Unknown], label="Model type", multiselect=False, value=None, interactive=True, ) with gr.Column(): precision = gr.Dropdown( choices=[i.value.name for i in utils.Precision if i != utils.Precision.Unknown], label="Precision", multiselect=False, value="float16", interactive=True, ) weight_type = gr.Dropdown( choices=[i.value.name for i in utils.WeightType], label="Weights type", multiselect=False, value="Original", interactive=True, ) base_model_name_textbox = gr.Textbox(label="Base model (for delta or adapter weights)") submit_button = gr.Button("Submit Eval") submission_result = gr.Markdown() submit_button.click( submit.add_new_eval, [ model_name_textbox, base_model_name_textbox, revision_name_textbox, precision, weight_type, model_type, ], submission_result, ) with gr.Row(): with gr.Accordion("📙 Citation", open=False): citation_button = gr.Textbox( value=about.CITATION_BUTTON_TEXT, label=about.CITATION_BUTTON_LABEL, lines=20, elem_id="citation-button", show_copy_button=True, ) scheduler = BackgroundScheduler() scheduler.add_job(restart_space, "interval", seconds=1800) scheduler.start() demo.queue(default_concurrency_limit=40).launch()