import json import math import subprocess import tempfile from pathlib import Path import numpy as np import pandas as pd import gradio as gr from huggingface_hub import snapshot_download from src.about import ( CITATION_BUTTON_LABEL, CITATION_BUTTON_TEXT, EVALUATION_QUEUE_TEXT, INTRODUCTION_TEXT, LLM_BENCHMARKS_TEXT, TITLE, Tasks, ) from src.display.css_html_js import custom_css from src.display.utils import ( BENCHMARK_COLS, COLS, EVAL_COLS, EVAL_TYPES, NUMERIC_INTERVALS, TYPES, AutoEvalColumn, ModelType, fields, WeightType, Precision ) from src.envs import API, DATA_PATH from src.populate import get_leaderboard_df def restart_space(): API.restart_space(repo_id=REPO_ID) raw_data, original_df = get_leaderboard_df(DATA_PATH, COLS, BENCHMARK_COLS) leaderboard_df = original_df.copy() def export_csv(df): csv_filename = Path(tempfile._get_default_tempdir()) / f"scandeval_leaderboard_{next(tempfile._get_candidate_names())}.csv" df = df.copy() df[AutoEvalColumn.model.name] = df[AutoEvalColumn.model.name].apply(lambda x: x.split(">")[1][:-3]) df.to_csv(csv_filename) return str(csv_filename) def plot_stats(data_path, plotting_library="plotly", columns=None, table=None): plots = {} files = Path(data_path).rglob("*.jsonl") models = None if table is not None: models = table.data[AutoEvalColumn.model.name].apply(lambda x: x.split(">")[1][:-3]).values.tolist() if columns is not None: scores = {(task.value.benchmark, task.value.metric): task.value.col_name for task in Tasks if task.value.col_name in columns} else: scores = {(task.value.benchmark, task.value.metric): task.value.col_name for task in Tasks} model_names = [] for file in files: with open(file) as f: for line in f: if not line.strip(): continue line = json.loads(line) if line["model"] not in model_names: model_names.append(line["model"]) if models is not None and line["model"] not in models: continue metrics = {} for r in line["results"]["raw"]["test"]: for k, v in r.items(): key = (line["dataset"], k) if key not in scores: continue val = plots.get(key, {}) val[line["model"]] = val.get(line["model"], []) + [v] plots[key] = val metrics.setdefault(k, []).append(v) # Boxplot # target_size = math.ceil(len(plots) ** 0.5) ncols = 2 # target_size if target_size ** 2 == len(plots) else target_size + 1 nrows = len(plots) // 2 + len(plots) % 2 # target_size if plotting_library == "matplotlib": import matplotlib.pyplot as plt if not plots: return plt.subplots(1, 1)[0] fig, axs = plt.subplots(nrows, ncols, figsize=(10 * ncols, 10 * nrows)) for i, (k, v) in enumerate(plots.items()): ax = axs[i // ncols, i % ncols] vk, vv = zip(*sorted(v.items())) ax.boxplot(vv, tick_labels=vk) # Tilt the x-axis labels slightly for tick in ax.get_xticklabels(): tick.set_rotation(5) ax.set_title(scores[k]) # fig.show() fig.tight_layout() # fig.savefig("results.png") elif plotting_library == "plotly": import plotly.express as px import plotly.graph_objects as go from plotly.subplots import make_subplots if not plots: return make_subplots(rows=1, cols=1) colors = dict(zip(model_names, px.colors.qualitative.Dark24)) plot_titles = [scores[k] for k in plots.keys()] fig = make_subplots(rows=nrows + 1, cols=ncols, subplot_titles=[AutoEvalColumn.average.name] + plot_titles, specs=[[{"colspan": ncols, "type": "scatterpolar"}] + [None] * (ncols - 1)] + [[{}] * ncols] * nrows, start_cell="top-left", vertical_spacing=0.05, horizontal_spacing=0.1) scatters = {} # print(fig.print_grid()) for i, (k, v) in enumerate(plots.items()): vk, vv = zip(*sorted(v.items())) for j, (label, data) in enumerate(zip(vk, vv)): # Adding a box trace for each label in the subplot # print(i // ncols + 2, i % ncols + 1) fig.add_trace(go.Box(y=data, name=label, boxpoints=False, marker_color=colors[label], showlegend=False, legendgroup=f'group-{label}'), row=i // ncols + 2, col=i % ncols + 1) if label not in scatters: scatters[label] = {} scatters[label][plot_titles[i]] = np.mean(data) if max(data) < 1 else np.mean(data) / 100.0 for label, data in scatters.items(): fig.add_trace(go.Scatterpolar( r=tuple(data.values()), theta=tuple(data.keys()), fill="toself", name=label, line_color=colors[label], legendgroup=f'group-{label}', ), row=1, col=1) # Update xaxis properties for each subplot to rotate and center labels for i in range(nrows * ncols): fig.update_xaxes(tickangle=-15, row=i // ncols + 2, col=i % ncols + 1) fig.update_layout( height=500 * (nrows + 1), width=700 * ncols, showlegend=True, # Prevent plot from getting its top cut off, not showing the titles # margin keyword has no effect, instead we do: title_yanchor="top", ) # fig.show() else: raise ValueError(f"Unknown plotting library: {plotting_library}") return fig # Searching and filtering def update_table_and_plot( 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) export_filename = export_csv(df) df = (df.style .format(precision=2, thousands=",", decimal=".") .highlight_max(props="background-color: lightgreen; color: black;", axis=0, subset=df.columns[1:]) .highlight_between(props="color: red;", axis=0, subset=df.columns[1:], left=-np.inf, right=-np.inf) ) fig = plot_stats(DATA_PATH, columns=df.data.columns, table=df) return df, fig, export_filename def search_table(df: pd.DataFrame, query: str) -> pd.DataFrame: if query.lower().startswith("not "): return df[~(df[AutoEvalColumn.model.name].str.contains(query[4:], case=False))] else: return df[(df[AutoEvalColumn.model.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] ] 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: if _q != "" and not _q.lower().startswith("not "): 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=[AutoEvalColumn.model.name, AutoEvalColumn.precision.name, AutoEvalColumn.revision.name] # ) filtered_df = filtered_df.drop_duplicates( subset=[AutoEvalColumn.model.name] ) for _q in queries: if _q != "" and _q.lower().startswith("not "): filtered_df = search_table(filtered_df, _q) 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 # filtered_df = df[df[AutoEvalColumn.still_on_hub.name] == True] # type_emoji = [t[0] for t in type_query] # filtered_df = filtered_df.loc[df[AutoEvalColumn.model_type_symbol.name].isin(type_emoji)] # filtered_df = filtered_df.loc[df[AutoEvalColumn.precision.name].isin(precision_query + ["None"])] # 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.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 fields(AutoEvalColumn) if not c.hidden and not c.never_hidden ], value=[ c.name for c in fields(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(visible=False): deleted_models_visibility = gr.Checkbox( value=False, label="Show gated/private/deleted models", interactive=True ) with gr.Column(min_width=320, visible=False): # with gr.Box(elem_id="box-filter"): filter_columns_type = gr.CheckboxGroup( label="Model types", choices=[t.to_str() for t in ModelType], value=[t.to_str() for t in ModelType], interactive=True, elem_id="filter-columns-type", ) filter_columns_precision = gr.CheckboxGroup( label="Precision", choices=[i.value.name for i in Precision], value=[i.value.name for i in Precision], interactive=True, elem_id="filter-columns-precision", ) filter_columns_size = gr.CheckboxGroup( label="Model sizes (in billions of parameters)", choices=list(NUMERIC_INTERVALS.keys()), value=list(NUMERIC_INTERVALS.keys()), interactive=True, elem_id="filter-columns-size", ) with gr.Tabs(elem_classes="tab-buttons") as tabs: with gr.TabItem("🏅 LLM Benchmark", elem_id="llm-benchmark-tab-table", id=0): cols_to_show = [c.name for c in fields(AutoEvalColumn) if c.never_hidden] + shown_columns.value leaderboard_table = gr.components.Dataframe( value=(leaderboard_df[cols_to_show].style .format(precision=2, thousands=",", decimal=".") .highlight_max(props="background-color: lightgreen; color: black;", axis=0, subset=cols_to_show[1:]) .highlight_between(props="color: red;", axis=0, subset=cols_to_show[1:], left=-np.inf, right=-np.inf) ), headers=[c.name for c in fields(AutoEvalColumn) if c.never_hidden] + shown_columns.value, datatype=TYPES, elem_id="leaderboard-table", interactive=False, visible=True, ) leaderboard_file = gr.File(interactive=False, value=export_csv(leaderboard_df[cols_to_show]), visible=True) with gr.TabItem("📊 LLM Plots", elem_id="llm-benchmark-tab-plot", id=1): leaderboard_plot = gr.components.Plot(plot_stats(DATA_PATH)) # Dummy leaderboard for handling the case when the user uses backspace key hidden_leaderboard_table_for_search = gr.components.Dataframe( value=original_df[COLS], headers=COLS, datatype=TYPES, visible=False, ) search_bar.submit( update_table_and_plot, [ hidden_leaderboard_table_for_search, shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, deleted_models_visibility, search_bar, ], [ leaderboard_table, leaderboard_plot, leaderboard_file, ], ) for selector in [shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, deleted_models_visibility]: selector.change( update_table_and_plot, [ hidden_leaderboard_table_for_search, shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, deleted_models_visibility, search_bar, ], [ leaderboard_table, leaderboard_plot, leaderboard_file, ], 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, # 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=EVAL_COLS, # datatype=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=EVAL_COLS, # datatype=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 ModelType if t != ModelType.Unknown], # label="Model type", # multiselect=False, # value=None, # interactive=True, # ) # with gr.Column(): # precision = gr.Dropdown( # choices=[i.value.name for i in Precision if i != Precision.Unknown], # label="Precision", # multiselect=False, # value="float16", # interactive=True, # ) # weight_type = gr.Dropdown( # choices=[i.value.name for i in 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( # 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=CITATION_BUTTON_TEXT, # label=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()