open-llm-leaderboard/open_llm_leaderboard
14k
1import os2import logging3import time4import gradio as gr5import datasets6from huggingface_hub import snapshot_download, WebhooksServer, WebhookPayload, RepoCard7from gradio_leaderboard import Leaderboard, ColumnFilter, SelectColumns8 9from src.display.about import (10 CITATION_BUTTON_LABEL,11 CITATION_BUTTON_TEXT,12 EVALUATION_QUEUE_TEXT,13 FAQ_TEXT,14 INTRODUCTION_TEXT,15 LLM_BENCHMARKS_TEXT,16 TITLE,17)18from src.display.css_html_js import custom_css19from src.display.utils import (20 BENCHMARK_COLS,21 COLS,22 EVAL_COLS,23 EVAL_TYPES,24 AutoEvalColumn,25 ModelType,26 Precision,27 WeightType,28 fields,29)30from src.envs import (31 API,32 EVAL_REQUESTS_PATH,33 AGGREGATED_REPO,34 HF_TOKEN,35 QUEUE_REPO,36 REPO_ID,37 HF_HOME,38)39from src.populate import get_evaluation_queue_df, get_leaderboard_df40from src.submission.submit import add_new_eval41from src.tools.plots import create_metric_plot_obj, create_plot_df, create_scores_df42 43# Configure logging44logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")45 46 47# Convert the environment variable "LEADERBOARD_FULL_INIT" to a boolean value, defaulting to True if the variable is not set.48# This controls whether a full initialization should be performed.49DO_FULL_INIT = os.getenv("LEADERBOARD_FULL_INIT", "True") == "True"50 51def restart_space():52 API.restart_space(repo_id=REPO_ID, token=HF_TOKEN)53 54 55def time_diff_wrapper(func):56 def wrapper(*args, **kwargs):57 start_time = time.time()58 result = func(*args, **kwargs)59 end_time = time.time()60 diff = end_time - start_time61 logging.info(f"Time taken for {func.__name__}: {diff} seconds")62 return result63 64 return wrapper65 66 67@time_diff_wrapper68def download_dataset(repo_id, local_dir, repo_type="dataset", max_attempts=3, backoff_factor=1.5):69 """Download dataset with exponential backoff retries."""70 attempt = 071 while attempt < max_attempts:72 try:73 logging.info(f"Downloading {repo_id} to {local_dir}")74 snapshot_download(75 repo_id=repo_id,76 local_dir=local_dir,77 repo_type=repo_type,78 tqdm_class=None,79 etag_timeout=30,80 max_workers=8,81 )82 logging.info("Download successful")83 return84 except Exception as e:85 wait_time = backoff_factor**attempt86 logging.error(f"Error downloading {repo_id}: {e}, retrying in {wait_time}s")87 time.sleep(wait_time)88 attempt += 189 raise Exception(f"Failed to download {repo_id} after {max_attempts} attempts")90 91def get_latest_data_leaderboard():92 leaderboard_dataset = datasets.load_dataset(93 AGGREGATED_REPO, 94 "default", 95 split="train", 96 cache_dir=HF_HOME, 97 download_mode=datasets.DownloadMode.REUSE_DATASET_IF_EXISTS, # Uses the cached dataset 98 verification_mode="no_checks"99 )100 101 leaderboard_df = get_leaderboard_df(102 leaderboard_dataset=leaderboard_dataset, 103 cols=COLS,104 benchmark_cols=BENCHMARK_COLS,105 )106 107 return leaderboard_df108 109def get_latest_data_queue():110 eval_queue_dfs = get_evaluation_queue_df(EVAL_REQUESTS_PATH, EVAL_COLS)111 return eval_queue_dfs112 113def init_space():114 """Initializes the application space, loading only necessary data."""115 if DO_FULL_INIT:116 # These downloads only occur on full initialization117 try:118 download_dataset(QUEUE_REPO, EVAL_REQUESTS_PATH)119 except Exception:120 restart_space()121 122 # Always redownload the leaderboard DataFrame123 leaderboard_df = get_latest_data_leaderboard()124 125 # Evaluation queue DataFrame retrieval is independent of initialization detail level126 eval_queue_dfs = get_latest_data_queue()127 128 return leaderboard_df, eval_queue_dfs129 130 131# Calls the init_space function with the `full_init` parameter determined by the `do_full_init` variable.132# This initializes various DataFrames used throughout the application, with the level of initialization detail controlled by the `do_full_init` flag.133leaderboard_df, eval_queue_dfs = init_space()134finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df = eval_queue_dfs135 136 137# Data processing for plots now only on demand in the respective Gradio tab138def load_and_create_plots():139 plot_df = create_plot_df(create_scores_df(leaderboard_df))140 return plot_df141 142def init_leaderboard(dataframe):143 return Leaderboard(144 value = dataframe,145 datatype=[c.type for c in fields(AutoEvalColumn)],146 select_columns=SelectColumns(147 default_selection=[c.name for c in fields(AutoEvalColumn) if c.displayed_by_default],148 cant_deselect=[c.name for c in fields(AutoEvalColumn) if c.never_hidden or c.dummy],149 label="Select Columns to Display:",150 ),151 search_columns=[AutoEvalColumn.model.name, AutoEvalColumn.fullname.name, AutoEvalColumn.license.name],152 hide_columns=[c.name for c in fields(AutoEvalColumn) if c.hidden],153 filter_columns=[154 ColumnFilter(AutoEvalColumn.model_type.name, type="checkboxgroup", label="Model types"),155 ColumnFilter(AutoEvalColumn.precision.name, type="checkboxgroup", label="Precision"),156 ColumnFilter(157 AutoEvalColumn.params.name,158 type="slider",159 min=0.01,160 max=150,161 label="Select the number of parameters (B)",162 ),163 ColumnFilter(164 AutoEvalColumn.still_on_hub.name, type="boolean", label="Private or deleted", default=True165 ),166 ColumnFilter(167 AutoEvalColumn.merged.name, type="boolean", label="Contains a merge/moerge", default=True168 ),169 ColumnFilter(AutoEvalColumn.moe.name, type="boolean", label="MoE", default=False),170 ColumnFilter(AutoEvalColumn.not_flagged.name, type="boolean", label="Flagged", default=True),171 ],172 bool_checkboxgroup_label="Hide models",173 )174 175 176demo = gr.Blocks(css=custom_css)177with demo:178 gr.HTML(TITLE)179 gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")180 181 with gr.Tabs(elem_classes="tab-buttons") as tabs:182 with gr.TabItem("🏅 LLM Benchmark", elem_id="llm-benchmark-tab-table", id=0):183 leaderboard = init_leaderboard(leaderboard_df)184 185 with gr.TabItem("📈 Metrics through time", elem_id="llm-benchmark-tab-table", id=2):186 with gr.Row():187 with gr.Column():188 plot_df = load_and_create_plots()189 chart = create_metric_plot_obj(190 plot_df,191 [AutoEvalColumn.average.name],192 title="Average of Top Scores and Human Baseline Over Time (from last update)",193 )194 gr.Plot(value=chart, min_width=500)195 with gr.Column():196 plot_df = load_and_create_plots()197 chart = create_metric_plot_obj(198 plot_df,199 BENCHMARK_COLS,200 title="Top Scores and Human Baseline Over Time (from last update)",201 )202 gr.Plot(value=chart, min_width=500)203 204 with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=3):205 gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")206 207 with gr.TabItem("❗FAQ", elem_id="llm-benchmark-tab-table", id=4):208 gr.Markdown(FAQ_TEXT, elem_classes="markdown-text")209 210 with gr.TabItem("🚀 Submit ", elem_id="llm-benchmark-tab-table", id=5):211 with gr.Column():212 with gr.Row():213 gr.Markdown(EVALUATION_QUEUE_TEXT, elem_classes="markdown-text")214 215 with gr.Row():216 gr.Markdown("# ✉️✨ Submit your model here!", elem_classes="markdown-text")217 218 with gr.Row():219 with gr.Column():220 model_name_textbox = gr.Textbox(label="Model name")221 revision_name_textbox = gr.Textbox(label="Revision commit", placeholder="main")222 model_type = gr.Dropdown(223 choices=[t.to_str(" : ") for t in ModelType if t != ModelType.Unknown],224 label="Model type",225 multiselect=False,226 value=ModelType.FT.to_str(" : "),227 interactive=True,228 )229 230 with gr.Column():231 precision = gr.Dropdown(232 choices=[i.value.name for i in Precision if i != Precision.Unknown],233 label="Precision",234 multiselect=False,235 value="float16",236 interactive=True,237 )238 weight_type = gr.Dropdown(239 choices=[i.value.name for i in WeightType],240 label="Weights type",241 multiselect=False,242 value="Original",243 interactive=True,244 )245 base_model_name_textbox = gr.Textbox(label="Base model (for delta or adapter weights)")246 247 with gr.Column():248 with gr.Accordion(249 f"✅ Finished Evaluations ({len(finished_eval_queue_df)})",250 open=False,251 ):252 with gr.Row():253 finished_eval_table = gr.components.Dataframe(254 value=finished_eval_queue_df,255 headers=EVAL_COLS,256 datatype=EVAL_TYPES,257 row_count=5,258 )259 with gr.Accordion(260 f"🔄 Running Evaluation Queue ({len(running_eval_queue_df)})",261 open=False,262 ):263 with gr.Row():264 running_eval_table = gr.components.Dataframe(265 value=running_eval_queue_df,266 headers=EVAL_COLS,267 datatype=EVAL_TYPES,268 row_count=5,269 )270 271 with gr.Accordion(272 f"⏳ Pending Evaluation Queue ({len(pending_eval_queue_df)})",273 open=False,274 ):275 with gr.Row():276 pending_eval_table = gr.components.Dataframe(277 value=pending_eval_queue_df,278 headers=EVAL_COLS,279 datatype=EVAL_TYPES,280 row_count=5,281 )282 283 submit_button = gr.Button("Submit Eval")284 submission_result = gr.Markdown()285 submit_button.click(286 add_new_eval,287 [288 model_name_textbox,289 base_model_name_textbox,290 revision_name_textbox,291 precision,292 weight_type,293 model_type,294 ],295 submission_result,296 )297 298 with gr.Row():299 with gr.Accordion("📙 Citation", open=False):300 citation_button = gr.Textbox(301 value=CITATION_BUTTON_TEXT,302 label=CITATION_BUTTON_LABEL,303 lines=20,304 elem_id="citation-button",305 show_copy_button=True,306 )307 308 demo.load(fn=get_latest_data_leaderboard, inputs=None, outputs=[leaderboard])309 demo.load(fn=get_latest_data_queue, inputs=None, outputs=[finished_eval_table, running_eval_table, pending_eval_table])310 311demo.queue(default_concurrency_limit=40)312 313# Start ephemeral Spaces on PRs (see config in README.md)314from gradio_space_ci.webhook import IS_EPHEMERAL_SPACE, SPACE_ID, configure_space_ci315 316def enable_space_ci_and_return_server(ui: gr.Blocks) -> WebhooksServer:317 # Taken from https://huggingface.co/spaces/Wauplin/gradio-space-ci/blob/075119aee75ab5e7150bf0814eec91c83482e790/src/gradio_space_ci/webhook.py#L61318 # Compared to original, this one do not monkeypatch Gradio which allows us to define more webhooks.319 # ht to Lucain!320 if SPACE_ID is None:321 print("Not in a Space: Space CI disabled.")322 return WebhooksServer(ui=demo)323 324 if IS_EPHEMERAL_SPACE:325 print("In an ephemeral Space: Space CI disabled.")326 return WebhooksServer(ui=demo)327 328 card = RepoCard.load(repo_id_or_path=SPACE_ID, repo_type="space")329 config = card.data.get("space_ci", {})330 print(f"Enabling Space CI with config from README: {config}")331 332 return configure_space_ci(333 blocks=ui,334 trusted_authors=config.get("trusted_authors"),335 private=config.get("private", "auto"),336 variables=config.get("variables", "auto"),337 secrets=config.get("secrets"),338 hardware=config.get("hardware"),339 storage=config.get("storage"),340 )341 342# Create webhooks server (with CI url if in Space and not ephemeral)343webhooks_server = enable_space_ci_and_return_server(ui=demo)344 345# Add webhooks346@webhooks_server.add_webhook347async def update_leaderboard(payload: WebhookPayload) -> None:348 """Redownloads the leaderboard dataset each time it updates"""349 if payload.repo.type == "dataset" and payload.event.action == "update":350 datasets.load_dataset(351 AGGREGATED_REPO, 352 "default", 353 split="train", 354 cache_dir=HF_HOME, 355 download_mode=datasets.DownloadMode.FORCE_REDOWNLOAD, 356 verification_mode="no_checks"357 )358 359@webhooks_server.add_webhook 360async def update_queue(payload: WebhookPayload) -> None:361 """Redownloads the queue dataset each time it updates"""362 if payload.repo.type == "dataset" and payload.event.action == "update":363 download_dataset(QUEUE_REPO, EVAL_REQUESTS_PATH)364 365webhooks_server.launch()366