from functools import reduce import json import pickle import os import re import gradio as gr import pandas as pd from tqdm.autonotebook import tqdm from utils.model_size import get_model_parameters_memory from refresh import TASK_TO_METRIC, TASKS, PRETTY_NAMES, TASKS_CONFIG, BOARDS_CONFIG, load_results from envs import REPO_ID from refresh import PROPRIETARY_MODELS, SENTENCE_TRANSFORMERS_COMPATIBLE_MODELS, CROSS_ENCODERS, BI_ENCODERS, TASK_DESCRIPTIONS, EXTERNAL_MODEL_TO_LINK, make_clickable_model PROPRIETARY_MODELS = { make_clickable_model(model, link=EXTERNAL_MODEL_TO_LINK.get(model, f"https://huggingface.co/spaces/{REPO_ID}")) for model in PROPRIETARY_MODELS } SENTENCE_TRANSFORMERS_COMPATIBLE_MODELS = { make_clickable_model(model, link=EXTERNAL_MODEL_TO_LINK.get(model, f"https://huggingface.co/spaces/{REPO_ID}")) for model in SENTENCE_TRANSFORMERS_COMPATIBLE_MODELS } CROSS_ENCODERS = { make_clickable_model(model, link=EXTERNAL_MODEL_TO_LINK.get(model, f"https://huggingface.co/spaces/{REPO_ID}")) for model in CROSS_ENCODERS } BI_ENCODERS = { make_clickable_model(model, link=EXTERNAL_MODEL_TO_LINK.get(model, f"https://huggingface.co/spaces/{REPO_ID}")) for model in BI_ENCODERS } def make_datasets_clickable(df): """Does not work""" if "BornholmBitextMining" in df.columns: link = "https://huggingface.co/datasets/strombergnlp/bornholmsk_parallel" df = df.rename( columns={f'BornholmBitextMining': 'BornholmBitextMining',}) return df # 1. Force headers to wrap # 2. Force model column (maximum) width # 3. Prevent model column from overflowing, scroll instead # 4. Prevent checkbox groups from taking up too much space css = """ table > thead { white-space: normal } table { --cell-width-1: 250px } table > tbody > tr > td:nth-child(2) > div { overflow-x: auto } .filter-checkbox-group { max-width: max-content; } """ """ Each inner tab can have the following keys: - language: The language of the leaderboard - language_long: [optional] The long form of the language - description: The description of the leaderboard - credits: [optional] The credits for the leaderboard - data: The data for the leaderboard """ # No more refreshing manually, happens daily # def get_refresh_function(task_category, task_list): # def _refresh(): # data_task_category = get_mteb_data(tasks=[task_category], datasets=task_list) # data_task_category.drop(columns=["Embedding Dimensions", "Max Tokens"], inplace=True) # return data_task_category # return _refresh # def get_refresh_overall_function(tasks): # return lambda: get_mteb_average(tasks)[0] # load in the pre-calculated `all_data_tasks` and `boards_data` print(f"Loading pre-calculated data....") all_data_tasks = load_results("all_data_tasks") boards_data = load_results("boards_data") #### Caclulate Metadata # Exact, add all non-nan integer values for every dataset NUM_SCORES = 0 DATASETS = [] MODELS = [] # LANGUAGES = [] for d in all_data_tasks: # NUM_SCORES += d.iloc[:, 1:].apply(lambda x: sum([1 for y in x if isinstance(y, float) and not np.isnan(y)]), axis=1).sum() cols_to_ignore = 4 if "Average" in d.columns else 3 # Count number of scores including only non-nan floats & excluding the rank column NUM_SCORES += d.iloc[:, cols_to_ignore:].notna().sum().sum() # Exclude rank & model name column (first two); Do not count different language versions as different datasets DATASETS += [i.split(" ")[0] for i in d.columns[cols_to_ignore:]] # LANGUAGES += [i.split(" ")[-1] for i in d.columns[cols_to_ignore:]] MODELS += d["Model"].tolist() NUM_DATASETS = len(set(DATASETS)) # NUM_LANGUAGES = len(set(LANGUAGES)) NUM_MODELS = len(set(MODELS)) data = { "Overall": {"metric": "Various, refer to task tabs", "data": []} } for task in TASKS: data[task] = {"metric": TASKS_CONFIG[task]["metric_description"], "data": []} for board, board_config in BOARDS_CONFIG.items(): init_name = board_config["title"] if init_name in PRETTY_NAMES: init_name = PRETTY_NAMES[init_name] board_pretty_name = f"{init_name} leaderboard" acronym = board_config.get("acronym", None) board_icon = board_config.get("icon", None) if board_icon is None: board_icon = "" credits = board_config.get("credits", None) metric = board_config.get("metric", None) if board_config["has_overall"]: overall_pretty_name = board_pretty_name if acronym is not None: overall_pretty_name += f" ({board_config['acronym']})" data["Overall"]["data"].append({ "language": board_config["title"], "language_long": board_config["language_long"], "description": f"**Overall MTEB {overall_pretty_name}** 🔮{board_icon}", "data": boards_data[board]["data_overall"], # "refresh": get_refresh_overall_function(board_config["tasks"]), "credits": credits, "metric": metric, }) for task_category, task_category_list in board_config["tasks"].items(): task_icon = TASKS_CONFIG[task_category]['icon'] if "special_icons" in board_config and isinstance(board_config["special_icons"], dict): task_icon = board_config["special_icons"].get(task_category, task_icon) data[task_category]["data"].append({ "language": board_config["title"], "language_long": board_config["language_long"], "description": f"**{task_category} {board_pretty_name}** {task_icon}{board_icon}", "data": boards_data[board]["data_tasks"][task_category], # "refresh": get_refresh_function(task_category, task_category_list), "credits": credits, "metric": metric, }) dataframes = [] full_dataframes = [] tabs = [] # The following JavaScript function updates the URL parameters based on the selected task and language # Additionally, `update_url_task` and `update_url_language` are used to update the current task and language # The current task and language are stored in the `current_task_language` and `language_per_task` JSON objects # This is all a bit hacky, but it might be the only way to pass options to a JavaScript function via Gradio set_window_url_params = """ function(goalUrlObject) { const params = new URLSearchParams(window.location.search); for (const [key, value] of Object.entries(goalUrlObject)) { params.set(key, value); }; const queryString = '?' + params.toString(); console.log(queryString); window.history.replaceState({}, '', queryString); return []; } """ def update_url_task(event: gr.SelectData, current_task_language: dict, language_per_task: dict): current_task_language["task"] = event.target.id # Either use the cached language for this task or the 1st language try: current_task_language["language"] = language_per_task.get(event.target.id, event.target.children[1].children[0].id) except Exception as e: # is Overall tab, no description current_task_language["language"] = language_per_task.get(event.target.id, event.target.children[0].children[0].id) return current_task_language, language_per_task def update_url_language(event: gr.SelectData, current_task_language: dict, language_per_task: dict): current_task_language["language"] = event.target.id if "task" not in current_task_language: current_task_language["task"] = "overall" language_per_task[current_task_language["task"]] = event.target.id return current_task_language, language_per_task NUMERIC_INTERVALS = { "<100M": pd.Interval(0, 100, closed="right"), "100M to 250M": pd.Interval(100, 250, closed="right"), "250M to 500M": pd.Interval(250, 500, closed="right"), "500M to 1B": pd.Interval(500, 1000, closed="right"), ">1B": pd.Interval(1000, 1_000_000, closed="right"), } MODEL_TYPES = [ "Open", "Proprietary", "Sentence Transformers", "Cross-Encoders", "Bi-Encoders" ] def filter_data(search_query, model_types, model_sizes, *full_dataframes): output_dataframes = [] for df in full_dataframes: # Apply the search query if search_query: names = df["Model"].map(lambda x: re.match("(.+)", x).group(1)) masks = [] for query in search_query.split(";"): masks.append(names.str.lower().str.contains(query.lower())) df = df[reduce(lambda a, b: a | b, masks)] # Apply the model type filtering if set(model_types) != set(MODEL_TYPES): masks = [] for model_type in model_types: if model_type == "Open": masks.append(~df["Model"].isin(PROPRIETARY_MODELS)) elif model_type == "Proprietary": masks.append(df["Model"].isin(PROPRIETARY_MODELS)) elif model_type == "Sentence Transformers": masks.append(df["Model"].isin(SENTENCE_TRANSFORMERS_COMPATIBLE_MODELS)) elif model_type == "Cross-Encoders": masks.append(df["Model"].isin(CROSS_ENCODERS)) elif model_type == "Bi-Encoders": masks.append(df["Model"].isin(BI_ENCODERS)) if masks: df = df[reduce(lambda a, b: a | b, masks)] else: df = pd.DataFrame(columns=df.columns) # Apply the model size filtering if set(model_sizes) != set(NUMERIC_INTERVALS.keys()): numeric_interval = pd.IntervalIndex(sorted([NUMERIC_INTERVALS[model_size] for model_size in model_sizes])) sizes = df["Model Size (Million Parameters)"].replace('', 0) mask = sizes.apply(lambda size: any(numeric_interval.contains(size))) df = df[mask] output_dataframes.append(df) return output_dataframes with gr.Blocks(css=css) as block: # Store the current task and language for updating the URL. This is a bit hacky, but it works # for passing the current task and language to the JavaScript function via Gradio current_task_language = gr.JSON(value=dict(), visible=False) language_per_task = gr.JSON(value=dict(), visible=False) gr.Markdown(f""" Massive Text Embedding Benchmark (MTEB) Leaderboard. To submit, refer to the MTEB GitHub repository 🤗 Refer to the [MTEB paper](https://arxiv.org/abs/2210.07316) for details on metrics, tasks and models. """) with gr.Row(): search_bar = gr.Textbox( label="Search Bar (separate multiple queries with `;`)", placeholder=" 🔍 Search for a model and press enter...", ) filter_model_type = gr.CheckboxGroup( label="Model types", choices=MODEL_TYPES, value=MODEL_TYPES, interactive=True, elem_classes=["filter-checkbox-group"] ) filter_model_sizes = gr.CheckboxGroup( label="Model sizes (in number of parameters)", choices=list(NUMERIC_INTERVALS.keys()), value=list(NUMERIC_INTERVALS.keys()), interactive=True, elem_classes=["filter-checkbox-group"], scale=2, ) with gr.Tabs() as outer_tabs: # Store the tabs for updating them on load based on URL parameters tabs.append(outer_tabs) for task, task_values in data.items(): metric = task_values["metric"] task_tab_id = task.lower().replace(" ", "-") # Overall, Bitext Mining, Classification, etc. pretty_task_name = task if task not in PRETTY_NAMES.keys() else PRETTY_NAMES[task] with gr.Tab(pretty_task_name, id=task_tab_id) as task_tab: # For updating the 'task' in the URL task_tab.select(update_url_task, [current_task_language, language_per_task], [current_task_language, language_per_task]).then(None, [current_task_language], [], js=set_window_url_params) if "Overall" != task: gr.Markdown(TASK_DESCRIPTIONS[task]) with gr.Tabs() as task_tabs: # Store the task tabs for updating them on load based on URL parameters tabs.append(task_tabs) for item in task_values["data"]: item_tab_id = item["language"].lower().replace(" ", "-") # English, Chinese, French, etc. with gr.Tab(item["language"], id=item_tab_id) as item_tab: # For updating the 'language' in the URL item_tab.select(update_url_language, [current_task_language, language_per_task], [current_task_language, language_per_task], trigger_mode="always_last").then(None, [current_task_language], [], js=set_window_url_params) specific_metric = metric if item.get("metric", None) is not None: specific_metric = item['metric'] with gr.Row(): gr.Markdown(f""" {item['description']} - **Metric:** {specific_metric} - **Languages:** {item['language_long'] if 'language_long' in item else item['language']} {"- **Credits:** " + item['credits'] if ("credits" in item and item["credits"] is not None) else ''} """) with gr.Row(): datatype = ["number", "markdown"] + ["number"] * len(item["data"]) dataframe = gr.Dataframe(item["data"], datatype=datatype, type="pandas", height=500) dataframes.append(dataframe) full_dataframe = gr.Dataframe(item["data"], datatype=datatype, type="pandas", visible=False) full_dataframes.append(full_dataframe) # with gr.Row(): # refresh_button = gr.Button("Refresh") # refresh_button.click(item["refresh"], inputs=None, outputs=dataframe, concurrency_limit=20) gr.Markdown(f""" - **Total Datasets**: {NUM_DATASETS} - **Total Languages**: 113 - **Total Scores**: {NUM_SCORES} - **Total Models**: {NUM_MODELS} """ + r""" Made with ❤️ for NLP. If this work is useful to you, please consider citing: ```bibtex @article{muennighoff2022mteb, doi = {10.48550/ARXIV.2210.07316}, url = {https://arxiv.org/abs/2210.07316}, author = {Muennighoff, Niklas and Tazi, Nouamane and Magne, Lo{\"\i}c and Reimers, Nils}, title = {MTEB: Massive Text Embedding Benchmark}, publisher = {arXiv}, journal={arXiv preprint arXiv:2210.07316}, year = {2022} } ``` """) def set_tabs_on_load(request: gr.Request): """Set the selected tab based on the URL parameters on load.""" global tabs valid_task_keys = [child.id for child in tabs[0].children] return_tabs = [gr.Tabs()] * len(tabs) query_params = request.request.query_params task_key = query_params.get("task", "overall") if task_key not in valid_task_keys: task_key = "overall" return_tabs[0] = gr.Tabs(selected=task_key) tabs_idx = valid_task_keys.index(task_key) + 1 language_key = query_params.get("language", "english") return_tabs[tabs_idx] = gr.Tabs(selected=language_key) current_task_language = {"task": task_key, "language": language_key} language_per_task = {task_key: language_key} return return_tabs + [current_task_language, language_per_task] block.load(set_tabs_on_load, inputs=[], outputs=tabs + [current_task_language, language_per_task]) search_bar.submit(filter_data, inputs=[search_bar, filter_model_type, filter_model_sizes] + full_dataframes, outputs=dataframes) filter_model_type.change(filter_data, inputs=[search_bar, filter_model_type, filter_model_sizes] + full_dataframes, outputs=dataframes) filter_model_sizes.change(filter_data, inputs=[search_bar, filter_model_type, filter_model_sizes] + full_dataframes, outputs=dataframes) block.queue(max_size=10) block.launch() # Possible changes: # Could add graphs / other visual content # Could add verification marks # Sources: # https://huggingface.co/spaces/gradio/leaderboard # https://huggingface.co/spaces/huggingface-projects/Deep-Reinforcement-Learning-Leaderboard # https://getemoji.com/