import json import os from datetime import datetime, timezone import re from distutils.util import strtobool import gradio as gr import pandas as pd from apscheduler.schedulers.background import BackgroundScheduler from huggingface_hub import HfApi, snapshot_download from src.assets.css_html_js import custom_css, get_window_url_params from src.assets.text_content import ( CITATION_BUTTON_LABEL, CITATION_BUTTON_TEXT, EVALUATION_QUEUE_TEXT, INTRODUCTION_TEXT, LLM_BENCHMARKS_TEXT, TITLE, BOTTOM_LOGO, ) from src.display_models.get_model_metadata import DO_NOT_SUBMIT_MODELS, ModelType from src.display_models.utils import ( AutoEvalColumn, EvalQueueColumn, fields, styled_error, styled_message, styled_warning, ) from src.load_from_hub import get_all_requested_models, get_evaluation_queue_df, get_leaderboard_df, is_model_on_hub from src.rate_limiting import user_submission_permission pd.set_option("display.precision", 1) # clone / pull the lmeh eval data H4_TOKEN = os.environ.get("H4_TOKEN", None) QUEUE_REPO = "open-ko-llm-leaderboard/requests" RESULTS_REPO = "open-ko-llm-leaderboard/results" PRIVATE_QUEUE_REPO = "open-ko-llm-leaderboard/private-requests" PRIVATE_RESULTS_REPO = "open-ko-llm-leaderboard/private-results" IS_PUBLIC = bool(strtobool(os.environ.get("IS_PUBLIC", "True"))) EVAL_REQUESTS_PATH = "eval-queue" EVAL_RESULTS_PATH = "eval-results" EVAL_REQUESTS_PATH_PRIVATE = "eval-queue-private" EVAL_RESULTS_PATH_PRIVATE = "eval-results-private" api = HfApi(token=H4_TOKEN) def restart_space(): api.restart_space(repo_id="upstage/open-ko-llm-leaderboard", token=H4_TOKEN) # Rate limit variables RATE_LIMIT_PERIOD = 7 RATE_LIMIT_QUOTA = 5 # Column selection COLS = [c.name for c in fields(AutoEvalColumn) if not c.hidden] TYPES = [c.type for c in fields(AutoEvalColumn) if not c.hidden] COLS_LITE = [c.name for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden] TYPES_LITE = [c.type for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden] if not IS_PUBLIC: COLS.insert(2, AutoEvalColumn.precision.name) TYPES.insert(2, AutoEvalColumn.precision.type) EVAL_COLS = [c.name for c in fields(EvalQueueColumn)] EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)] BENCHMARK_COLS = [ c.name for c in [ AutoEvalColumn.arc, AutoEvalColumn.hellaswag, AutoEvalColumn.mmlu, AutoEvalColumn.truthfulqa, AutoEvalColumn.commongen_v2, # TODO: Uncomment when we have results for these # AutoEvalColumn.ethicalverification, ] ] snapshot_download(repo_id=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH, repo_type="dataset", tqdm_class=None) snapshot_download(repo_id=RESULTS_REPO, local_dir=EVAL_RESULTS_PATH, repo_type="dataset", tqdm_class=None) requested_models, users_to_submission_dates = get_all_requested_models(EVAL_REQUESTS_PATH) original_df = get_leaderboard_df(EVAL_RESULTS_PATH, COLS, BENCHMARK_COLS) leaderboard_df = original_df.copy() models = original_df["model_name_for_query"].tolist() # needed for model backlinks in their to the leaderboard # Commented out because it causes infinite restart loops in local # to_be_dumped = f"models = {repr(models)}\n" # with open("models_backlinks.py", "w") as f: # f.write(to_be_dumped) # print(to_be_dumped) ( finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df, failed_eval_queue_df, ) = get_evaluation_queue_df(EVAL_REQUESTS_PATH, EVAL_COLS) ## INTERACTION FUNCTIONS def add_new_eval( model: str, base_model: str, revision: str, precision: str, private: bool, weight_type: str, model_type: str, ): precision = precision.split(" ")[0] current_time = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ") num_models_submitted_in_period = user_submission_permission(model, users_to_submission_dates, RATE_LIMIT_PERIOD) if num_models_submitted_in_period > RATE_LIMIT_QUOTA: error_msg = f"Organisation or user `{model.split('/')[0]}`" error_msg += f"already has {num_models_submitted_in_period} model requests submitted to the leaderboard " error_msg += f"in the last {RATE_LIMIT_PERIOD} days.\n" error_msg += "Please wait a couple of days before resubmitting, so that everybody can enjoy using the leaderboard 🤗" return styled_error(error_msg) if model_type is None or model_type == "": return styled_error("Please select a model type.") # check the model actually exists before adding the eval if revision == "": revision = "main" if weight_type in ["Delta", "Adapter"]: base_model_on_hub, error = is_model_on_hub(base_model, revision) if not base_model_on_hub: return styled_error(f'Base model "{base_model}" {error}') if not weight_type == "Adapter": model_on_hub, error = is_model_on_hub(model, revision) if not model_on_hub: return styled_error(f'Model "{model}" {error}') model_info = api.model_info(repo_id=model, revision=revision) size_pattern = re.compile(r"(\d+\.)?\d+(b|m)") try: model_size = round(model_info.safetensors["total"] / 1e9, 3) except AttributeError: try: size_match = re.search(size_pattern, model.lower()) model_size = size_match.group(0) model_size = round(float(model_size[:-1]) if model_size[-1] == "b" else float(model_size[:-1]) / 1e3, 3) except AttributeError: return 65 size_factor = 8 if (precision == "GPTQ" or "GPTQ" in model) else 1 model_size = size_factor * model_size try: license = model_info.cardData["license"] except Exception: license = "?" print("adding new eval") eval_entry = { "model": model, "base_model": base_model, "revision": revision, "private": private, "precision": precision, "weight_type": weight_type, "status": "PENDING", "submitted_time": current_time, "model_type": model_type, } user_name = "" model_path = model if "/" in model: user_name = model.split("/")[0] model_path = model.split("/")[1] if user_name == "upstage": return styled_message( "The model participating as a Host in Upstage does not conduct evaluations to ensure the transparency and fairness of the leaderboard. Please take this into consideration." ) OUT_DIR = f"{EVAL_REQUESTS_PATH}/{user_name}" os.makedirs(OUT_DIR, exist_ok=True) out_path = f"{OUT_DIR}/{model_path}_eval_request_{private}_{precision}_{weight_type}.json" # Check if the model has been forbidden: if out_path.split("eval-queue/")[1] in DO_NOT_SUBMIT_MODELS: return styled_warning("Model authors have requested that their model be not submitted on the leaderboard.") # Check for duplicate submission if f"{model}_{revision}_{precision}" in requested_models: return styled_warning("This model has been already submitted.") with open(out_path, "w") as f: f.write(json.dumps(eval_entry)) api.upload_file( path_or_fileobj=out_path, path_in_repo=out_path.split("eval-queue/")[1], repo_id=QUEUE_REPO, repo_type="dataset", commit_message=f"Add {model} to eval queue", ) # remove the local file os.remove(out_path) return styled_message( "Your request has been submitted to the evaluation queue!\nPlease wait for up to an hour for the model to show in the PENDING list." ) # Basics def change_tab(query_param: str): query_param = query_param.replace("'", '"') query_param = json.loads(query_param) if isinstance(query_param, dict) and "tab" in query_param and query_param["tab"] == "evaluation": return gr.Tabs.update(selected=1) else: return gr.Tabs.update(selected=0) # Searching and filtering def update_table(hidden_df: pd.DataFrame, current_columns_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) if query != "": filtered_df = search_table(filtered_df, query) df = select_columns(filtered_df, columns) return df def search_table(df: pd.DataFrame, query: str) -> pd.DataFrame: return df[(df[AutoEvalColumn.dummy.name].str.contains(query, case=False))] def select_columns(df: pd.DataFrame, columns: list) -> pd.DataFrame: always_here_cols = [ AutoEvalColumn.model_type_symbol.name, AutoEvalColumn.model.name, ] # We use COLS to maintain sorting filtered_df = df[ always_here_cols + [c for c in COLS if c in df.columns and c in columns] + [AutoEvalColumn.dummy.name] ] return filtered_df NUMERIC_INTERVALS = { "Unknown": pd.Interval(-1, 0, closed="right"), "0~3B": pd.Interval(0, 3, closed="right"), "3~7B": pd.Interval(3, 7.3, closed="right"), "7~13B": pd.Interval(7.3, 13, closed="right"), "13~35B": pd.Interval(13, 35, closed="right"), "35~60B": pd.Interval(35, 60, closed="right"), "60B+": pd.Interval(60, 10000, closed="right"), } 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[AutoEvalColumn.still_on_hub.name] == True] type_emoji = [t[0] for t in type_query] filtered_df = filtered_df[df[AutoEvalColumn.model_type_symbol.name].isin(type_emoji)] filtered_df = filtered_df[df[AutoEvalColumn.precision.name].isin(precision_query)] numeric_interval = pd.IntervalIndex(sorted([NUMERIC_INTERVALS[s] for s in size_query])) params_column = pd.to_numeric(df[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(TITLE) gr.Markdown(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 and press ENTER...", show_label=False, elem_id="search-bar", ) with gr.Row(): shown_columns = gr.CheckboxGroup( choices=[ c for c in COLS if c not in [ AutoEvalColumn.dummy.name, AutoEvalColumn.model.name, AutoEvalColumn.model_type_symbol.name, AutoEvalColumn.still_on_hub.name, ] ], value=[ c for c in COLS_LITE if c not in [ AutoEvalColumn.dummy.name, AutoEvalColumn.model.name, AutoEvalColumn.model_type_symbol.name, AutoEvalColumn.still_on_hub.name, ] ], label="Select columns to show", elem_id="column-select", interactive=True, ) with gr.Row(): deleted_models_visibility = gr.Checkbox( value=True, 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=[ ModelType.PT.to_str(), # ModelType.FT.to_str(), ModelType.IFT.to_str(), ModelType.RL.to_str(), ], value=[ ModelType.PT.to_str(), # ModelType.FT.to_str(), ModelType.IFT.to_str(), ModelType.RL.to_str(), ], interactive=True, elem_id="filter-columns-type", ) filter_columns_precision = gr.CheckboxGroup( label="Precision", choices=["torch.float16"], #, "torch.bfloat16", "torch.float32", "8bit", "4bit", "GPTQ"], value=["torch.float16"], #, "torch.bfloat16", "torch.float32", "8bit", "4bit", "GPTQ"], interactive=False, elem_id="filter-columns-precision", ) filter_columns_size = gr.CheckboxGroup( label="Model sizes", choices=list(NUMERIC_INTERVALS.keys()), value=list(NUMERIC_INTERVALS.keys()), interactive=True, elem_id="filter-columns-size", ) leaderboard_table = gr.components.Dataframe( value=leaderboard_df[ [AutoEvalColumn.model_type_symbol.name, AutoEvalColumn.model.name] + shown_columns.value + [AutoEvalColumn.dummy.name] ], headers=[ AutoEvalColumn.model_type_symbol.name, AutoEvalColumn.model.name, ] + shown_columns.value + [AutoEvalColumn.dummy.name], datatype=TYPES, max_rows=None, elem_id="leaderboard-table", interactive=False, visible=True, ) # Dummy leaderboard for handling the case when the user uses backspace key hidden_leaderboard_table_for_search = gr.components.Dataframe( value=original_df, headers=COLS, datatype=TYPES, max_rows=None, visible=False, ) search_bar.submit( update_table, [ hidden_leaderboard_table_for_search, leaderboard_table, shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, deleted_models_visibility, search_bar, ], leaderboard_table, ) shown_columns.change( update_table, [ hidden_leaderboard_table_for_search, leaderboard_table, shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, deleted_models_visibility, search_bar, ], leaderboard_table, queue=True, ) filter_columns_type.change( update_table, [ hidden_leaderboard_table_for_search, leaderboard_table, shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, deleted_models_visibility, search_bar, ], leaderboard_table, queue=True, ) filter_columns_precision.change( update_table, [ hidden_leaderboard_table_for_search, leaderboard_table, shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, deleted_models_visibility, search_bar, ], leaderboard_table, queue=True, ) filter_columns_size.change( update_table, [ hidden_leaderboard_table_for_search, leaderboard_table, shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, deleted_models_visibility, search_bar, ], leaderboard_table, queue=True, ) deleted_models_visibility.change( update_table, [ hidden_leaderboard_table_for_search, leaderboard_table, 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(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(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=EVAL_COLS, datatype=EVAL_TYPES, max_rows=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=EVAL_COLS, datatype=EVAL_TYPES, max_rows=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=EVAL_COLS, datatype=EVAL_TYPES, max_rows=5, ) with gr.Accordion( f"❌ Failed Evaluations ({len(failed_eval_queue_df)})", open=False, ): with gr.Row(): pending_eval_table = gr.components.Dataframe( value=failed_eval_queue_df, headers=EVAL_COLS, datatype=EVAL_TYPES, max_rows=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", placeholder="main") private = gr.Checkbox(False, label="Private", visible=not IS_PUBLIC) model_type = gr.Dropdown( choices=[ ModelType.PT.to_str(" : "), # ModelType.FT.to_str(" : "), ModelType.IFT.to_str(" : "), ModelType.RL.to_str(" : "), ], label="Model type", multiselect=False, value=None, interactive=True, ) with gr.Column(): precision = gr.Dropdown( choices=[ "float16", # "bfloat16", # "8bit (LLM.int8)", # "4bit (QLoRA / FP4)", # "GPTQ" ], label="Precision", multiselect=False, value="float16", interactive=True, ) weight_type = gr.Dropdown( choices=["Original", "Delta", "Adapter"], 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 Evalulation!") submission_result = gr.Markdown() submit_button.click( add_new_eval, [ model_name_textbox, base_model_name_textbox, revision_name_textbox, precision, private, weight_type, model_type, ], submission_result, ) with gr.Row(): with gr.Accordion("📙 Citation", open=False): citation_button = gr.Textbox( value=CITATION_BUTTON_TEXT, label=CITATION_BUTTON_LABEL, elem_id="citation-button", ).style(show_copy_button=True) gr.HTML(BOTTOM_LOGO) dummy = gr.Textbox(visible=False) demo.load( change_tab, dummy, tabs, _js=get_window_url_params, ) scheduler = BackgroundScheduler() scheduler.add_job(restart_space, "interval", seconds=1800) scheduler.start() demo.queue(concurrency_count=40).launch()