new template test

app.py

@@ -0,0 +1,128 @@
+import gradio as gr
+
+from src.assets.text_content import TITLE, INTRODUCTION_TEXT
+from src.utils import get_data, compare_plots, filter_search
+
+############################ For Leaderboards #############################
+DATA_PATH = 'versions'
+latest_flag = True #Set flag to iclude latest data in Details and Versions Tab
+latest_df, latest_vname, previous_df, previous_vname = get_data(DATA_PATH, latest_flag)
+
+global prev_df
+prev_df = previous_df[0]
+def select_prev_df(name):
+    ind = previous_vname.index(name)
+    prev_df = previous_df[ind]
+    return prev_df
+
+############################ For Plots ####################################
+global plot_df, MODEL_COLS
+plot_df = latest_df[0]
+MODEL_COLS = list(plot_df['Model'].unique())
+
+
+############# MAIN APPLICATION ######################
+demo = gr.Blocks()
+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("🥇 TOFU Leaderboard", elem_id="llm-benchmark-tab-table", id=0):
+            with gr.Row():
+                search_bar = gr.Textbox(
+                    placeholder=" 🔍 Search for models - separate multiple queries with `;` and press ENTER...",
+                    show_label=False,
+                    elem_id="search-bar",
+                )
+                        
+            leaderboard_table = gr.components.Dataframe(
+                value=latest_df[0],
+                elem_id="leaderboard-table",
+                interactive=False,
+                visible=True,
+            )
+
+            # Add a dummy leaderboard to handle search queries from the latest_df and not update latest_df
+            dummy_leaderboard_table = gr.components.Dataframe(
+                value=latest_df[0],
+                elem_id="leaderboard-table",
+                interactive=False,
+                visible=False,
+            )
+                
+            search_bar.submit(
+                filter_search,
+                [dummy_leaderboard_table, search_bar],
+                leaderboard_table,
+                queue=True
+            )
+        with gr.TabItem("📈 Plot", id=3):
+            with gr.Row():
+                model_cols = gr.CheckboxGroup(
+                    MODEL_COLS, 
+                    label="Select Models 🤖", 
+                    value=[],
+                    elem_id="column-select",
+                    interactive=True,
+                )
+
+            with gr.Row():
+                plot_grdf = gr.DataFrame(
+                    value=plot_df,
+                    visible=False
+                )
+            with gr.Row():
+                # Output block for the plot
+                plot_output = gr.Plot()
+
+            model_cols.change(
+                compare_plots,
+                [plot_grdf, model_cols],
+                plot_output,
+                queue=True
+            )
+
+        with gr.TabItem("🔄 Versions and Details", elem_id="details", id=2):
+            with gr.Row():
+                ver_selection = gr.Dropdown(
+                    previous_vname, label="Select Version 🕹️", value=previous_vname[0]
+                )
+            with gr.Row():
+                search_bar_prev = gr.Textbox(
+                    placeholder=" 🔍 Search for models - separate multiple queries with `;` and press ENTER...",
+                    show_label=False,
+                    elem_id="search-bar-2",
+                )
+
+            prev_table = gr.components.Dataframe(
+                value=prev_df,
+                elem_id="leaderboard-table",
+                interactive=False,
+                visible=True,
+            )
+
+            dummy_prev_table = gr.components.Dataframe(
+                value=prev_df,
+                elem_id="leaderboard-table",
+                interactive=False,
+                visible=False,
+            )
+
+            search_bar_prev.submit(
+                filter_search,
+                [dummy_prev_table, search_bar_prev],
+                prev_table,
+                queue=True
+            )
+
+            ver_selection.change(
+                select_prev_df,
+                [ver_selection],
+                prev_table,
+                queue=True
+            )
+
+    demo.load()
+demo.queue()
+demo.launch()

src/assets/text_content.py

@@ -0,0 +1,12 @@
+TITLE = """<h1 align="center" id="space-title"> 🏆 TOFU Leaderboard</h1>"""
+
+INTRODUCTION_TEXT = """
+TOFU leaderboard description.
+"""
+
+SHORT_NAMES = {
+    "KL": "KL",
+    "Grad Ascent": "Grad Ascent",
+    "Gradient Difference": "Grad Diff",
+    "Oracle": "Oracle",
+}

src/utils.py

@@ -0,0 +1,236 @@
+import os
+import pandas as pd
+import matplotlib.pyplot as plt
+import numpy as np
+
+from src.assets.text_content import SHORT_NAMES
+
+def update_cols(df: pd.DataFrame) -> pd.DataFrame:
+    '''
+    Change three header rows to a single header row
+    Args:
+        df: Raw dataframe containing 3 separate header rows
+            Remove this function if the dataframe has only one header row
+    Returns:
+        df: Updated dataframe which has only 1 header row instead of 3
+    '''
+    default_cols = list(df.columns)
+
+    # First 4 columns are initalised in 'update', Append additional columns for games Model, Clemscore, ALL(PLayed) and ALL(Main Score)
+    update = ['Model', 'Clemscore', 'Played', 'Quality Score']
+    game_metrics = default_cols[4:]
+
+    # Change columns Names for each Game
+    for i in range(len(game_metrics)):
+        if i%3 == 0:
+            game = game_metrics[i]
+            update.append(str(game).capitalize() + "(Played)")
+            update.append(str(game).capitalize() + "(Quality Score)") 
+            update.append(str(game).capitalize() + "(Quality Score[std])")
+
+    # Create a dict to change names of the columns
+    map_cols = {}
+    for i in range(len(default_cols)):
+        map_cols[default_cols[i]] = str(update[i])
+
+    df = df.rename(columns=map_cols)
+    df = df.iloc[2:]
+
+    return df
+
+def process_df(df: pd.DataFrame) -> pd.DataFrame:
+    '''
+    Process dataframe - Remove repition in model names, convert datatypes to sort by "float" instead of "str"
+    Args:
+        df: Unprocessed Dataframe (after using update_cols)
+    Returns:
+        df: Processed Dataframe
+    '''
+
+    # Change column type to float from str
+    list_column_names = list(df.columns)
+    model_col_name = list_column_names[0]
+    for col in list_column_names:
+        if col != model_col_name:
+            df[col] = df[col].astype(float)
+
+    # Remove repetition in model names, if any
+    models_list = []
+    for i in range(len(df)):
+        model_name = df.iloc[i][model_col_name]
+        splits = model_name.split('--')
+        splits = [split.replace('-t0.0', '') for split in splits] # Comment to not remove -t0.0
+        if splits[0] == splits[1]:
+            models_list.append(splits[0])
+        else:
+            models_list.append(splits[0] + "--" + splits[1])
+    df[model_col_name] = models_list
+    
+    return df
+
+def get_data(path: str, flag: bool):
+    '''
+    Get a list of all version names and respective Dataframes 
+    Args: 
+        path: Path to the directory containing CSVs of different versions -> v0.9.csv, v1.0.csv, ....
+        flag: Set this flag to include the latest version in Details and Versions tab
+    Returns: 
+        latest_df: singular list containing dataframe of the latest version of the leaderboard with only 4 columns 
+        latest_vname: list of the name of latest version 
+        previous_df: list of dataframes for previous versions (can skip latest version if required) 
+        previous_vname: list of the names for the previous versions (INCLUDED IN Details and Versions Tab)
+    '''
+    # Check if Directory is empty
+    list_versions = os.listdir(path)
+    if not list_versions:
+        print("Directory is empty")
+
+    else:
+        files = [file for file in list_versions if file.endswith('.csv')]
+        files.sort(reverse=True)
+        file_names = [os.path.splitext(file)[0] for file in files]
+
+        DFS = []
+        for file in files:
+            df = pd.read_csv(os.path.join(path, file))
+            df = update_cols(df) # Remove if by default there is only one header row
+            df = process_df(df) # Process Dataframe
+            df = df.sort_values(by=list(df.columns)[1], ascending=False) # Sort by clemscore
+            DFS.append(df)
+
+        # Only keep relavant columns for the main leaderboard
+        latest_df_dummy = DFS[0]
+        all_columns = list(latest_df_dummy.columns)
+        keep_columns = all_columns[0:4]
+        latest_df_dummy = latest_df_dummy.drop(columns=[c for c in all_columns if c not in keep_columns])
+
+        latest_df = [latest_df_dummy]
+        latest_vname = [file_names[0]]
+        previous_df = []
+        previous_vname = []
+        for df, name in zip(DFS, file_names):
+            previous_df.append(df)
+            previous_vname.append(name) 
+        
+        if not flag:
+            previous_df.pop(0)
+            previous_vname.pop(0)
+
+        return latest_df, latest_vname, previous_df, previous_vname
+    
+    return None
+
+
+# ['Model', 'Clemscore', 'All(Played)', 'All(Quality Score)']
+def compare_plots(df: pd.DataFrame, LIST: list):
+    '''
+    Quality Score v/s % Played plot by selecting models
+    Args:
+        LIST: The list of models to show in the plot, updated from frontend
+    Returns:
+        fig: The plot
+    '''
+    short_names = label_map(LIST)
+
+    list_columns = list(df.columns)
+    df = df[df[list_columns[0]].isin(LIST)]
+
+    X = df[list_columns[2]]
+    fig, ax = plt.subplots()
+    for model in LIST:
+        short = short_names[model]
+        # same_flag = short_names[model][1]
+        model_df = df[df[list_columns[0]] == model]
+        x = model_df[list_columns[2]]
+        y = model_df[list_columns[3]]
+        color = plt.cm.rainbow(x / max(X))  # Use a colormap for different colors
+        plt.scatter(x, y, color=color)
+        # if same_flag:
+        plt.annotate(f'{short}', (x, y), textcoords="offset points", xytext=(0, -15), ha='center', rotation=0)
+        # else:
+        #     plt.annotate(f'{short}', (x, y), textcoords="offset points", xytext=(20, -3), ha='center', rotation=0)
+    ax.grid(which='both', color='grey', linewidth=1, linestyle='-', alpha=0.2)
+    ax.set_xticks(np.arange(0,110,10))
+    plt.xlim(-10, 110)
+    plt.ylim(-10, 110)
+    plt.xlabel('% Played')
+    plt.ylabel('Quality Score')
+    plt.title('Overview of benchmark results')
+    plt.show()
+
+    return fig
+
+def shorten_model_name(full_name):
+    # Split the name into parts
+    parts = full_name.split('-')
+
+    # Process the name parts to keep only the parts with digits (model sizes and versions)
+    short_name_parts = [part for part in parts if any(char.isdigit() for char in part)]
+
+    if len(parts) == 1:
+        short_name = ''.join(full_name[0:min(3, len(full_name))])
+    else:
+        # Join the parts to form the short name
+        short_name = '-'.join(short_name_parts)
+
+        # Remove any leading or trailing hyphens
+        short_name = full_name[0] + '-'+ short_name.strip('-')
+
+    return short_name
+
+def label_map(model_list: list) -> dict:
+    '''
+    Generate a map from long names to short names, to plot them in frontend graph
+    Define the short names in src/assets/text_content.py
+    Args: 
+        model_list: A list of long model names
+    Returns:
+        short_name: A map from long to list of short name + indication if models are same or different
+    '''
+    short_names = {}
+    for model_name in model_list:
+        # splits = model_name.split('--')
+        # if len(splits) != 1:
+        #     splits[0] = SHORT_NAMES[splits[0] + '-']
+        #     splits[1] = SHORT_NAMES[splits[1] + '-']
+        #     # Define the short name and indicate there are two different models
+        #     short_names[model_name] = [splits[0] + '--' + splits[1], 0]
+        # else:
+        if model_name in SHORT_NAMES:
+            short_name = SHORT_NAMES[model_name]
+        else:
+            short_name = shorten_model_name(model_name)
+
+        # Define the short name and indicate both models are same
+        short_names[model_name] = short_name
+
+    return short_names
+
+def filter_search(df: pd.DataFrame, query: str) -> pd.DataFrame:
+    '''
+    Filter the dataframe based on the search query
+    Args:
+        df: Unfiltered dataframe
+        query: a string of queries separated by ";"
+    Return:
+        filtered_df: Dataframe containing searched queries in the 'Model' column 
+    '''
+    queries = query.split(';')
+    list_cols = list(df.columns)
+    df_len = len(df)
+    filtered_models = []
+    models_list = list(df[list_cols[0]])
+    for q in queries:
+        q = q.lower()
+        for i in range(df_len):
+            model_name = models_list[i]
+            if q in model_name.lower():
+                filtered_models.append(model_name) # Append model names containing query q
+
+    filtered_df = df[df[list_cols[0]].isin(filtered_models)]
+
+    if query == "":
+        return df
+
+    return filtered_df
+    

versions/v1.0.csv

@@ -0,0 +1,9 @@
+Method,Compute,PPL,Truth,ROUGE,MAPO
+DPO,Forget,0.0768753815825175,0.3986066518668836,0.9565206583678539,0.8712100761297452
+DPO,Retain,0.5996286412275281,0.3602207419553394,0.6659869365178336,0.10860875787554469
+Grad Ascent,Forget,0.01276005409910086,0.8739653533368998,0.023760115576687335,0.557119299008513
+Grad Ascent,Retain,0.9153365174437643,0.5117768071328227,0.13019174875223205,0.2736769500895253
+IDK,Forget,0.9575847571359651,0.14495165859171177,0.5215967278097287,0.9513180970650936
+IDK,Retain,0.5081723023409522,0.7260250131902866,0.46407442478973215,0.9008803129332287
+KL,Forget,0.23284021819861755,0.7973023013038227,0.9713336423092905,0.674807833567933
+KL,Retain,0.04023188230471908,0.7071738714102987,0.4663170982373773,0.6658539062921722