import gradio as gr import pandas as pd import matplotlib.pyplot as plt # Function to load data from a given CSV file def load_data(model,version,metrics): version = version.replace("%", "p") file_path = f'versions/{model}-{version}.csv' # Replace with your file paths df = pd.read_csv(file_path) # we only want specific columns and in a specific order # column_names : Method,Model,WD,Forget Rate,Epoch,LR,Compute,ROUGE Real Authors,ROUGE SEM Real Authors,Truth Ratio Real Authors,Truth Ratio SEM Real Authors,Prob. Real Authors,Prob. SEM Real Authors,ROUGE-P Real Authors,ROUGE-P SEM Real Authors,TTR Real Authors,TTR SEM Real Authors,ROUGE Real World,ROUGE SEM Real World,Truth Ratio Real World,Truth Ratio SEM Real World,Prob. Real World,Prob. SEM Real World,ROUGE-P Real World,ROUGE-P SEM Real World,TTR Real World,TTR SEM Real World,ROUGE Retain,ROUGE SEM Retain,Truth Ratio Retain,Truth Ratio SEM Retain,Prob. Retain,Prob. SEM Retain,ROUGE-P Retain,ROUGE-P SEM Retain,TTR Retain,TTR SEM Retain,KS Test Retain,Wilcoxon PVal Retain,Wilcoxon Stat Retain,ROUGE Forget,ROUGE SEM Forget,Truth Ratio Forget,Truth Ratio SEM Forget,Prob. Forget,Prob. SEM Forget,ROUGE-P Forget,ROUGE-P SEM Forget,TTR Forget,TTR SEM Forget,KS Test Forget,Wilcoxon PVal Forget,Wilcoxon Stat Forget,KS Test Real Authors,KS Test PVal Real Authors,Wilcoxon PVal Real Authors,Wilcoxon Stat Real Authors,KS Test Real World,KS Test PVal Real World,Wilcoxon PVal Real World,Wilcoxon Stat Real World,KS Test PVal Retain,KS Test PVal Forget,Model Utility,Forget Quality column_names = ["Method", "Model Utility", "Forget Quality", "ROUGE Real Authors", "Truth Ratio Real Authors", "Prob. Real Authors", "ROUGE Real World", "Truth Ratio Real World", "Prob. Real World", "ROUGE Retain", "Truth Ratio Retain", "Prob. Retain", "ROUGE Forget", "Truth Ratio Forget", "Prob. Forget", ] #based on the metrics, remove the columns that are not needed if "ROUGE" not in metrics: column_names = [x for x in column_names if "ROUGE" not in x] if "Truth Ratio" not in metrics: column_names = [x for x in column_names if "Truth Ratio" not in x] if "Prob." not in metrics: column_names = [x for x in column_names if "Prob." not in x] #if there is a column with name WD, modify each entry in Method to include WD: method (WD = wd) if "WD" in df.columns: #get the WD column entry for each row and add it to the method name df["Method"] = df["Method"] + " (WD = " + df["WD"].astype(str) + ")" df = df[column_names] # if there are multiple rows with the same method, keep only the one with the highest product of model utility and forget quality product = df["Model Utility"] * df["Forget Quality"] df["product"] = product df = df.sort_values(by="product", ascending=False) df = df.drop_duplicates(subset=["Method"], keep="first") df = df.drop(columns=["product"]) return df # def style_leaderboard(df): # make color red for background if column has "Forget" in it # Function for searching in the leaderboard def search_leaderboard(df, query): if query == "": return df else: return df[df['Method'].str.contains(query)] # Function to change the version of the leaderboard def change_version(model, version): new_df = load_data(model, version) return new_df # Function to create plots from plotter import create_plots # Initialize Gradio app demo = gr.Blocks() with demo: gr.Markdown(""" ## 🥇 TOFU Leaderboard The TOFU dataset is a benchmark designed to evaluate the unlearning performance of large language models in realistic scenarios. This unique dataset consists of question-answer pairs that are based on the autobiographies of 200 fictitious authors, entirely generated by the GPT-4 model. The primary objective of this task is to effectively unlearn a fine-tuned model using different portions of the forget set. """) with gr.Tabs(): with gr.TabItem("Leaderboard"): with gr.Row(): version_dropdown = gr.Dropdown( choices=["1%", "5%", "10%"], label="🔄 Select Forget Percentage", value="10%", ) model_dropdown = gr.Dropdown( choices=["llama", "phi"], label="🔄 Select Base Model", value="llama", ) metrics_checkbox = gr.CheckboxGroup( label="Select Metrics", choices=["ROUGE", "Truth Ratio", "Prob."], default=["ROUGE", "Truth Ratio", "Prob."], ) with gr.Row(): search_bar = gr.Textbox( placeholder="Search for methods...", show_label=False, ) leaderboard_table = gr.components.Dataframe( value=load_data("llama", "10%"), interactive=True, visible=True, ) version_dropdown.change( change_version, inputs=[model_dropdown,version_dropdown,metrics_checkbox], outputs=leaderboard_table ) model_dropdown.change( change_version, inputs=[model_dropdown,version_dropdown,metrics_checkbox], outputs=leaderboard_table ) search_bar.change( search_leaderboard, inputs=[leaderboard_table, search_bar,metrics_checkbox], outputs=leaderboard_table ) metrics_checkbox.change( change_version, inputs=[model_dropdown,version_dropdown,metrics_checkbox], outputs=leaderboard_table ) # # Dynamically update the choices for the methods checkbox # def update_method_choices(version): # df = load_data(version) # methods = df['Method'].unique() # methods_checkbox.update(choices=methods) # return df # version_dropdown_plots.change( # update_method_choices, # inputs=version_dropdown_plots, # outputs=[methods_checkbox, plot_output] # ) # methods_checkbox.change( # create_plots, # inputs=[methods_checkbox, leaderboard_table], # outputs=plot_output # ) # Launch the app gr.Markdown(""" ## Applicability 🚀 The dataset is in QA format, making it ideal for use with popular chat models such as Llama2, Mistral, or Qwen. However, it also works for any other large language model. The corresponding code base is written for the Llama2 model, but can be easily adapted to other models. ## Installation ``` conda create -n tofu python=3.10 conda activate tofu conda install pytorch pytorch-cuda=11.8 -c pytorch -c nvidia conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit pip install -r requirements.txt ``` ## Loading the Dataset To load the dataset, use the following code: ```python from datasets import load_dataset dataset = load_dataset("locuslab/TOFU","full") ``` ### Push to Leaderboard How to push your results to the leaderboard? """) demo.launch()