"""A gradio app that renders a static leaderboard. This is used for Hugging Face Space.""" import ast import argparse import glob import pickle import plotly import gradio as gr import numpy as np import pandas as pd import gradio as gr import pandas as pd from pathlib import Path import json from constants import * from datetime import datetime, timezone # from datasets import Dataset, load_dataset, concatenate_datasets import os, uuid from utils_display import model_info from constants import column_names, LEADERBOARD_REMARKS, DEFAULT_K, LEADERBOARD_REMARKS_MAIN import pytz from data_utils import post_processing, get_random_item # get the last updated time from the elo_ranks.all.jsonl file LAST_UPDATED = None # with open("_intro.md", "r") as f: # INTRO_MD = f.read() INTRO_MD = "" with open("_about_us.md", "r") as f: ABOUT_MD = f.read() with open("_header.md", "r") as f: HEADER_MD = f.read() with open("_metrics.md", "r") as f: METRICS_MD = f.read() raw_data = None original_df = None # available_models = [] # to be filled in later available_models = list(model_info.keys()) def df_filters(mode_selection_radio, show_open_source_model_only): global original_df # remove the rows when the model contains "❌" original_df = original_df[~original_df["Model"].str.contains("❌")] modes = { "greedy": ["greedy"], "sampling (Temp=0.5)": ["sampling"], "all": ["greedy", "sampling"] } # filter the df by the mode_selection_radio default_main_df = original_df[original_df["Mode"].isin(modes[mode_selection_radio])] default_main_df.insert(0, "", range(1, 1 + len(default_main_df))) return default_main_df.copy() def _gstr(text): return gr.Text(text, visible=False) def _tab_leaderboard(): global original_df, available_models # with gr.TabItem("📊 Main", elem_id="od-benchmark-tab-table-ablation", id=0, elem_classes="subtab"): if True: default_main_df = original_df.copy() # default_main_df.insert(0, "", range(1, 1 + len(default_main_df))) # default_main_df_no_task = default_main_df.copy() default_mode = "greedy" default_main_df = df_filters(default_mode, False) with gr.Row(): with gr.Column(scale=5): mode_selection_radio = gr.Radio(["greedy", "all"], show_label=False, elem_id="rank-column-radio", value=default_mode) # with gr.Row(): # with gr.Column(scale=2): leaderboard_table = gr.components.Dataframe( value=default_main_df, datatype= ["number", "markdown", "markdown", "number"], # max_rows=None, height=6000, elem_id="leaderboard-table", interactive=False, visible=True, column_widths=[50, 260, 100, 100, 120, 120, 100,100,110,100], wrap=True # min_width=60, ) # checkbox_show_task_categorized.change(fn=length_margin_change, inputs=[length_margin_choices, gr.Text("main", visible=False), checkbox_show_task_categorized, show_open_source_model_only, rank_column_radio], outputs=[leaderboard_table]) # show_open_source_model_only.change(fn=length_margin_change, inputs=[length_margin_choices, gr.Text("main", visible=False), checkbox_show_task_categorized, show_open_source_model_only, rank_column_radio], outputs=[leaderboard_table]) # rank_column_radio.change(fn=length_margin_change, inputs=[length_margin_choices, gr.Text("main", visible=False), checkbox_show_task_categorized, show_open_source_model_only, rank_column_radio], outputs=[leaderboard_table]) mode_selection_radio.change(fn=df_filters, inputs=[mode_selection_radio, _gstr("")], outputs=[leaderboard_table]) def sample_explore_item(model_name, size_H, size_W): # print(model_name, size_H, size_W) explore_item = get_random_item(model_name, size_H, size_W) if explore_item is None: return "No item found", "No item found", "No item found", "No item found" model_name = explore_item['Model'] example_id = explore_item['id'] puzzle_md = f"### 🦓 Puzzle [{example_id}]:\n\n" + explore_item['puzzle'].replace("## Clues:", "### **Clues:**").replace("\n", "
") model_reasoning_md = f"### 🤖 Reasoning of {model_name}:\n\n {explore_item['reasoning']}" model_prediction_md = f"### 💬 Answer of {model_name}:\n\n**Json format:** {str(explore_item['solution']).replace('___', 'null')}" + \ "\n\n**Table format:**\n" + explore_item['solution_table_md'] puzzle_solved = explore_item['correct_cells'] == explore_item['total_cells'] cell_acc = explore_item["correct_cells"] / explore_item["total_cells"] * 100 model_eval_md = f"### 🆚 Evaluation:\n\n **Total Cells**: {explore_item['total_cells']} | **Correct Cells**: {explore_item['correct_cells']} | **Puzzle solved**: {puzzle_solved} | **Cell Acc**: {cell_acc:.2f}%" turht_solution_md = f"### ✅ Truth Solution:\n\n{explore_item['truth_solution_table']}" return puzzle_md, model_reasoning_md, model_prediction_md, model_eval_md, turht_solution_md def _tab_explore(): global raw_data model_names = [item["Model"] for item in raw_data] # deduplicate and preserve the order model_names = list(dict.fromkeys(model_names)) with gr.Row(): model_selection = gr.Dropdown(choices = ["random"] + model_names, label="Model: ", elem_id="select-models", value="random", interactive=True) size_H_selection = gr.Dropdown(choices = ["random"] + [f"{i}" for i in range(2,7)], label="Num of Houses", elem_id="select-H", value="random", interactive=True) size_W_selection = gr.Dropdown(choices = ["random"] + [f"{i}" for i in range(2,7)], label="Num of Features", elem_id="select-W", value="random", interactive=True) with gr.Column(scale=1): # greedy_or_sample = gr.Radio(["greedy", "sampling"], show_label=False, elem_id="greedy-or-sample", value="greedy", interactive=True) gr.Markdown("### 🚀 Click below to sample a puzzle. ⬇️ ") explore_button = gr.Button("🦓 Sample a Zebra Puzzle!", elem_id="explore-button") puzzle_md = gr.Markdown("### 🦓 Puzzle: \n\nTo be loaded", elem_id="puzzle-md", elem_classes="box_md") model_reasoning_md = gr.Markdown("### 🤖 Reasoning: \n\nTo be loaded", elem_id="model-reasoning-md", elem_classes="box_md") model_prediction_md = gr.Markdown("### 💬 Answer: \n\nTo be loaded", elem_id="model-prediction-md", elem_classes="box_md") turht_solution_md = gr.Markdown("### ✅ Truth Solution: \n\nTo be loaded", elem_id="truth-solution-md", elem_classes="box_md") model_eval_md = gr.Markdown("### 🆚 Evaluation: \n\nTo be loaded", elem_id="model-eval-md", elem_classes="box_md") explore_button.click(fn=sample_explore_item, inputs=[model_selection, size_H_selection, size_W_selection], outputs=[puzzle_md, model_reasoning_md, model_prediction_md, model_eval_md, turht_solution_md]) def _tab_submit(): markdown_text = """ Please create an issue on our [Github](https://github.com/WildEval/ZeroEval/) repository to talk about your model. Then, we can test it for you and report the results here on the Leaderboard. If you would like to do local testing, please read our code [here](https://github.com/WildEval/ZeroEval/blob/main/src/evaluation/zebra_grid_eval.py) and apply for the access for the [private dataset](https://huggingface.co/datasets/allenai/ZebraLogicBench-private) that contains the truth solutions. """ gr.Markdown("## 🚀 Submit Your Results\n\n" + markdown_text, elem_classes="markdown-text") def build_demo(): global original_df, available_models, gpt4t_dfs, haiku_dfs, llama_dfs with gr.Blocks(theme=gr.themes.Soft(), css=css, js=js_light) as demo: gr.HTML(BANNER, elem_id="banner") # convert LAST_UPDATED to the PDT time LAST_UPDATED = datetime.now(pytz.timezone('US/Pacific')).strftime("%Y-%m-%d %H:%M:%S") header_md_text = HEADER_MD.replace("{LAST_UPDATED}", str(LAST_UPDATED)) gr.Markdown(header_md_text, elem_classes="markdown-text") with gr.Tabs(elem_classes="tab-buttons") as tabs: with gr.TabItem("🏅 Leaderboard", elem_id="od-benchmark-tab-table", id=0): _tab_leaderboard() with gr.TabItem("🔍 Explore", elem_id="od-benchmark-tab-table", id=1): _tab_explore() with gr.TabItem("🚀 Submit Your Results", elem_id="od-benchmark-tab-table", id=3): _tab_submit() with gr.TabItem("📮 About Us", elem_id="od-benchmark-tab-table", id=4): gr.Markdown(ABOUT_MD, elem_classes="markdown-text") with gr.Row(): with gr.Accordion("📙 Citation", open=False, elem_classes="accordion-label"): gr.Textbox( value=CITATION_TEXT, lines=7, label="Copy the BibTeX snippet to cite this source", elem_id="citation-button", show_copy_button=True) # ).style(show_copy_button=True) return demo def data_load(result_file): global raw_data, original_df print(f"Loading {result_file}") column_names_main = column_names.copy() # column_names_main.update({}) main_ordered_columns = ORDERED_COLUMN_NAMES # filter the data with Total Puzzles == 1000 click_url = True # read json file from the result_file with open(result_file, "r") as f: raw_data = json.load(f) # floatify the data, if possible for d in raw_data: for k, v in d.items(): try: d[k] = float(v) except: pass original_df = pd.DataFrame(raw_data) original_df = original_df[original_df["Total Puzzles"] == 1000] original_df = post_processing(original_df, column_names_main, ordered_columns=main_ordered_columns, click_url=click_url, rank_column=RANKING_COLUMN) # print(original_df.columns) if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument("--share", action="store_true") parser.add_argument("--result_file", help="Path to results table", default="ZeroEval-main/result_dirs/zebra-grid.summary.json") args = parser.parse_args() data_load(args.result_file) print(original_df) demo = build_demo() demo.launch(share=args.share, height=3000, width="100%")