Merge pull request #64 from yeonir/yeonirhee/sotopia-task

sotopia_space/benchmark.py

@@ -1,42 +1,13 @@
 import gradio as gr # type: ignore
 import pandas as pd
 from sotopia_space.constants import MODEL_OPTIONS
-from sotopia_space.utils import estimated_win_rate, make_clickable_model, styled_error, styled_warning, styled_message,apply_length_penalty 
+from sotopia_space.utils import post_processing
 
 LP_MODE = "v2"
 original_df, ablation_df = None, None
 LP_original_dfs = {} 
 DEFAULT_LP = 0.5 
 
-available_models = [] # to be filled in later
-original_df, ablation_df = None, None
-
-def slider_change_main(length_penalty):
-    global original_df, ablation_df, LP_MODE
-    adjusted_df = apply_length_penalty(original_df, ablation_df, length_penalty, mode=LP_MODE, LP_original_dfs=LP_original_dfs) 
-    adjusted_df = adjusted_df[["Model", "Overall Elo", "Task-Avg Elo", "# battles", "Length"]]
-    adjusted_df = adjusted_df.sort_values(by="Overall Elo", ascending=False)
-    # adjusted_df = add_winrates(adjusted_df, LP=length_penalty) 
-    # adjusted_df = adjusted_df.drop(columns=["Length"])
-    adjusted_df.insert(0, "Rank", range(1, 1 + len(adjusted_df)))
-    return adjusted_df
-
-def slider_change_full(length_penalty, show_winrate):
-    global original_df, ablation_df, LP_MODE
-    adjusted_df = apply_length_penalty(original_df, ablation_df, length_penalty, mode=LP_MODE, LP_original_dfs=LP_original_dfs)
-    # sort the model by the "Task-Avg Elo" column
-    adjusted_df = adjusted_df.sort_values(by="Overall Elo", ascending=False)
-    adjusted_df.drop(columns=["Overall Elo", "Task-Avg Elo", "# battles", "Length"], inplace=True)
-    if show_winrate == "none":
-        adjusted_df.insert(0, "Rank", range(1, 1 + len(adjusted_df)))
-        return adjusted_df
-    elif show_winrate == "gpt-3.5":
-        adjusted_df = add_winrates_tasks(adjusted_df, ref="gpt-3.5", LP=length_penalty)
-    elif show_winrate == "gpt-4":
-        adjusted_df = add_winrates_tasks(adjusted_df, ref="gpt-4", LP=length_penalty)
-    adjusted_df.insert(0, "Rank", range(1, 1 + len(adjusted_df)))
-    return adjusted_df
-
 def benchmark_table():
     global original_df, ablation_df
     global LP_original_dfs, LP_MODE
@@ -44,18 +15,15 @@ def benchmark_table():
     gr.Markdown(f"**Version**: sotopia (v1.01; 2024.04.22) | **# Examples**: 7200 | **# Models**: {len(MODEL_OPTIONS)} | **# Comparisons**: x", elem_classes="markdown-text")
                 
     with gr.TabItem("Vs GPT-3.5", elem_id="od-benchmark-tab-table-ablation", id=0, elem_classes="subtab"):
-        # original_df, ablation_df = skip_empty_original_df, skip_empty_ablation_df
-        original_df = pd.read_json('data_dir/models_vs_gpt35.jsonl', lines=True)
-        default_main_df = apply_length_penalty(original_df, ablation_df, length_penalty=DEFAULT_LP, mode=LP_MODE, LP_original_dfs=LP_original_dfs) 
+        default_main_df = pd.read_json('data_dir/models_vs_gpt35.jsonl', lines=True)
         default_main_df = default_main_df.sort_values(by="GOAL [0, 10]", ascending=False)
+        default_main_df = post_processing(default_main_df, None)
         # add a Rank column to the first columnn (starting from 1)
         default_main_df.insert(0, "Rank", range(1, 1 + len(default_main_df)))
+        
         with gr.Row():
             with gr.Column(scale=4):
-                gr.Markdown("**Vs GPT3.5**: The interlocutors are compared against GPT-3.5, the baseline model.") 
-            with gr.Column(scale=1):
-                length_penlty_slider = gr.Slider(minimum=0.1, maximum=1, step=0.1, value=DEFAULT_LP, label="Length Penalty", elem_id="length-penalty-slider") 
-        # checkbox_skip_empty = gr.Checkbox(label="Skip empty results", value=False, elem_id="skip-empty-checkbox", scale=2)
+                gr.Markdown("<h3>**Vs GPT3.5**: The interlocutors are compared against GPT-3.5, the baseline model.")
         TYPES = ["number", "markdown", "number"]
         leaderboard_table = gr.components.Dataframe(
             value=default_main_df,
@@ -66,5 +34,4 @@ def benchmark_table():
             interactive=False,
             visible=True,
             min_width=60,
-            ) 
-        #length_penlty_slider.change(fn=slider_change_main, inputs=[length_penlty_slider], outputs=[leaderboard_table])
+            )

sotopia_space/chat.py

@@ -91,8 +91,6 @@ def chat_introduction():
             🗄️ **Disclaimer**: User prompts and generated replies from the model may be collected solely for the purpose of pure academic research. By using this demo, users implicitly agree to these terms.
             """
         )
-    # with gr.Column(scale=1):
-    #     toggle_dark = gr.Button(value="Toggle Dark")
 
 def create_user_agent_dropdown(environment_id):
     _, environment_dict, agent_dict, relationship_dict = get_sotopia_profiles()

sotopia_space/constants.py

@@ -14,26 +14,8 @@ MODEL_OPTIONS = [
 ]
 
 MODEL_INFO = {
-    "Llama-2-13b-chat-hf.nosp": {"pretty_name": "Llama-2-13B-chat", "hf_model_id": "meta-llama/Llama-2-13b-chat-hf"},
-    "Llama-2-70b-chat-hf.nosp": {"pretty_name": "Llama-2-70B-chat", "hf_model_id": "meta-llama/Llama-2-70b-chat-hf"},
-    "Llama-2-7b-chat-hf.nosp": {"pretty_name": "Llama-2-7B-chat", "hf_model_id": "meta-llama/Llama-2-7b-chat-hf"},
-    "Llama-2-7b-chat-hf": {"pretty_name": "Llama-2-7B-chat (+sys prmpt)", "hf_model_id": "meta-llama/Llama-2-7b-chat-hf"},
-    "Mistral-7B-Instruct-v0.1": {"pretty_name": "Mistral-7B-Instruct", "hf_model_id": "mistralai/Mistral-7B-Instruct-v0.1"},
-    "Mistral-7B-Instruct-v0.2": {"pretty_name": "Mistral-7B-Instruct (v0.2)", "hf_model_id": "mistralai/Mistral-7B-Instruct-v0.2"},
-    "Mixtral-8x7B-Instruct-v0.1": {"pretty_name": "Mixtral-8x7B-Instruct", "hf_model_id": "mistralai/Mixtral-8x7B-Instruct-v0.1"},
-    "Nous-Hermes-2-Mixtral-8x7B-DPO": {"pretty_name": "Nous-Hermes-2-Mixtral-8x7B-DPO", "hf_model_id": "NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO"},
-    "Yi-34B-Chat": {"pretty_name": "Yi-34B-Chat", "hf_model_id": "01-ai/Yi-34B"},
-    "gemini-1.0-pro": {"pretty_name": "gemini-1.0-pro", "hf_model_id": "https://blog.google/technology/ai/google-gemini-ai/"},
-    "gemma-7b-it": {"pretty_name": "Gemma-7B-it", "hf_model_id": "google/gemma-7b"},
-    "gpt-3.5-turbo-0125": {"pretty_name": "gpt-3.5-turbo-0125", "hf_model_id": "https://platform.openai.com/"},
-    "gpt-4-0125-preview": {"pretty_name": "gpt-4-0125-preview", "hf_model_id": "https://platform.openai.com/"},
-    "tulu-2-dpo-70b": {"pretty_name": "Tulu-2-dpo-70b", "hf_model_id": "cmu-lti/tulu-2-dpo-70b"},
-    "vicuna-13b-v1.5": {"pretty_name": "Vicuna-13b-v1.5", "hf_model_id": "lmsys/vicuna-13b-v1.5"},
-    "zephyr-7b-beta": {"pretty_name": "Zephyr-7b-beta", "hf_model_id": "HuggingFaceH4/zephyr-7b-beta"},
-    "mistral-large-2402": {"pretty_name": "Mistral-Large", "hf_model_id": "https://mistral.ai/news/mistral-large/"},
-    "claude-3-opus-20240229": {"pretty_name": "Claude 3 Opus", "hf_model_id": "https://www.anthropic.com/claude"},
-    "claude-3-sonnet-20240229": {"pretty_name": "Claude 3 Sonnet", "hf_model_id": "https://www.anthropic.com/claude"},
-    "zephyr-7b-gemma-v0.1": {"pretty_name": "Zephyr-7b-Gemma", "hf_model_id": "HuggingFaceH4/zephyr-7b-gemma-v0.1"},
-    "Starling-LM-7B-beta": {"pretty_name": "StarlingLM-7B-beta", "hf_model_id": "Nexusflow/Starling-LM-7B-beta"},
-    "dbrx-instruct": {"pretty_name": "DBRX Instruct", "hf_model_id": "databricks/dbrx-instruct"}
+    "GPT-4": {"pretty_name": "GPT-4", "hf_model_id": "https://openai.com/blog/new-models-and-developer-products-announced-at-devday"},
+    "GPT-3.5": {"pretty_name": "GPT-3.5", "hf_model_id": "https://openai.com/blog/new-models-and-developer-products-announced-at-devday"},
+    "Llama-2": {"pretty_name": "Llama-2", "hf_model_id": "https://llama.meta.com/llama2/"},
+    "MPT": {"pretty_name": "MPT", "hf_model_id": "https://huggingface.co/docs/transformers/main/en/model_doc/mpt"}
 }

sotopia_space/utils.py

@@ -1,17 +1,6 @@
-from datasets import load_dataset, Dataset
-import os 
-import json 
-from datasets import load_dataset
-from datasets.utils.logging import disable_progress_bar # type: ignore
-from ui_constants import column_names, all_task_types
-import random 
-disable_progress_bar()
-import math 
+from ui_constants import column_names
 from sotopia_space.constants import MODEL_INFO
 
-id_to_data = None 
-model_len_info = None 
-
 
 def make_clickable_model(model_name):
     global MODEL_INFO
@@ -25,199 +14,26 @@ def make_clickable_model(model_name):
     else:
         return model_name
     
-
-def styled_error(error):
-    return f"<p style='color: red; font-size: 20px; text-align: center;'>{error}</p>"
-
-def styled_warning(warn):
-    return f"<p style='color: orange; font-size: 20px; text-align: center;'>{warn}</p>"
-
-def styled_message(message):
-    return f"<p style='color: green; font-size: 20px; text-align: center;'>{message}</p>"
-        
-
-def estimated_win_rate(elo_a, elo_b, LP=0):
-    """
-    Calculate the estimated win rate for player A against player B using their Elo ratings.
-    :param elo_a: Elo rating of player A
-    :param elo_b: Elo rating of player B
-    :return: Estimated win rate for player A
-    """
-    exponent = (elo_b - elo_a)*(10**LP) / 400
-    probability_a_wins = 1 / (1 + 10 ** exponent)
-    return (1-probability_a_wins)*100
-
-
-
 # Formats the columns
 def formatter(x):
     if type(x) is str:
         x = x
     else: 
-        x = round(x, 1)
+        x = round(x, 2)
     return x
 
-
-def add_winrates(current_df, LP=0):
-    df = current_df.copy()
-    elo_column = "Task-Avg Elo" 
-
-    # Correct way to filter the DataFrame and get the Elo rating for "gpt-4-0125-preview"
-    model_a_elo = df[df["Model"].str.contains("gpt-4")][elo_column].iloc[0]
-
-    # Correct way to filter the DataFrame and get the Elo rating for "gpt-3.5-turbo-0125"
-    model_b_elo = df[df["Model"].str.contains("gpt-3.5")][elo_column].iloc[0]
-
-    
-    # Calculate the win rate of "gpt-4-0125-preview" against all models
-    df['Win% vs GPT-4'] = df[elo_column].apply(lambda x: estimated_win_rate(model_a_elo, x, LP=LP)).apply(formatter)    
-    df['Win% vs GPT-3.5T'] = df[elo_column].apply(lambda x: estimated_win_rate(model_b_elo, x, LP=LP)).apply(formatter)    
-    # apply the formatter for the two new columns 
-    cols = list(df.columns)
-    cols.remove("# battles"); cols.append("# battles")
-    cols.remove("Length"); cols.append("Length")
-    df = df[cols]
-    return df
-
-def add_winrates_tasks(current_df, ref="gpt-4", LP=0):
-    new_df = current_df.copy()
-    for t in all_task_types:
-        column = column_names[t]
-        model_a_elo = current_df[current_df["Model"].str.contains(ref)][column].iloc[0]
-        new_df[column] = current_df[column].apply(lambda x: estimated_win_rate(model_a_elo, x, LP=LP)).apply(formatter)
-    return new_df
-        
-
 def post_processing(df, model_len_info):
     if model_len_info:
-        df["Length"] = df["model name "].apply(lambda x: model_len_info[x]["avg_len"])
+        df["Length"] = df["model_name"].apply(lambda x: model_len_info[x]["avg_len"])
 
     for col in df.columns:
-        if col == "model name ":
+        if col == "model_name":
             df[col] = df[col].apply(lambda x: x.replace(x, make_clickable_model(x)))
         else:
             df[col] = df[col].apply(formatter) # For numerical values 
     df.rename(columns=column_names, inplace=True)
-    df.sort_values(by="Task-Avg Elo", inplace=True, ascending=False)
+    df.sort_values(by="GOAL [0, 10]", inplace=True, ascending=False)
     # put the "Overall Elo" and "Task-Avg Elo" column to the front
     # add the length info
-    df = df[["Model", "Task-Avg Elo"] + [col for col in df.columns if col not in ["Model", "Task-Avg Elo"]]]
+    df = df[["model_name", "GOAL [0, 10]"] + [col for col in df.columns if col not in ["model_name", "GOAL [0, 10]"]]]
     return df
-
-def apply_length_penalty(original_df, ablation_df, length_penalty=0.2, mode='v1', LP_original_dfs=None):
-    """
-    Temporarily disable the length penalty feature
-    if mode == 'v2' and LP_original_dfs is not None:
-        L = f"{length_penalty:.1f}"
-        return LP_original_dfs[L]
-    original_df = original_df.copy()
-    ablation_df = ablation_df.copy()
-    # replace all values in original_df with the values as z = x - y * length_penalty where y is from ablation_df at the same row and column
-    # except for the "Model" column and the "# battles" column 
-    # do not assume the order of the rows are the same in both dataframes
-    for i, row in original_df.iterrows():
-        for col in original_df.columns:
-            if col == "Model" or col == "# battles" or col == "Length":
-                continue
-            # assert that the model names are the same in both dataframes
-            assert original_df.at[i, "Model"] == ablation_df[ablation_df["Model"] == row["Model"]]["Model"].values[0]
-            original_df[col] = original_df[col].astype(float)
-            if mode == "v1":
-                original_df.at[i, col] = original_df.at[i, col] - ablation_df[ablation_df["Model"] == row["Model"]][col].values[0] * length_penalty
-            elif mode == "v1.1":
-                diff = original_df.at[i, col] - ablation_df[ablation_df["Model"] == row["Model"]][col].values[0] 
-                original_df.at[i, col] = original_df.at[i, col] * (1-length_penalty) + diff*length_penalty
-    # post_processing
-    original_df = post_processing(original_df, model_len_info=None)
-    """
-    return original_df 
-
-def load_benchdata():
-    print("Loading sotopia data...")
-    bench_data = load_dataset("cmu-lti/sotopia", split="test")
-    return bench_data
-
-def load_benchdata_dict():
-    print("Loading sotopia data....")
-    bench_data = load_dataset("cmu-lti/sotopia", data_files="sotopia_episodes_v1_hf.jsonl")['train']
-    id_to_data = {}
-    for item in bench_data:
-        id_to_data[item["session_id"]] = item
-    return id_to_data
-
-def load_eval_results():
-    print("Loading sotopia Evaluation data...")
-    eval_results = load_dataset("WildEval/sotopia-Evaluation", "all", split="train")
-    return eval_results
-
-def load_infer_results(model_name):
-    print(f"Loading sotopia Results for {model_name}...")
-    infer_results = load_dataset("WildEval/sotopia-Results", model_name, split="train")
-    return infer_results
-
-def sample_an_eval_result(eval_results, model_list=[], tag_list=[]):
-    global id_to_data          
-    eval_results = list(eval_results)
-    random.shuffle(eval_results)
-    for eval_item in eval_results:  
-        # print(json.dumps(eval_item, indent=2))
-        # print(f"## Session ID: {eval_item['session_id']}")
-        # eval_item["eval_id"]
-        assignment = eval_item['assignment']
-        model_1, model_2 = eval_item['model_1'], eval_item['model_2']
-        model_A = model_1 if assignment['A'] == model_1 else model_2
-        model_B = model_2 if assignment['B'] == model_2 else model_1
-        if len(model_list) >= 2:
-            if model_A not in model_list or model_B not in model_list:
-                continue
-        elif len(model_list) == 1:
-            if model_A != model_list[0] and model_B != model_list[0]:
-                continue
-        else:
-            pass 
-        if tag_list:
-            if set(tag_list).isdisjoint(set(eval_item['tags'])):
-                continue
-        winner = eval_item['winner']
-        # print(f"## Model A: {model_A} | Model B: {model_B} | Winner: {winner}")
-        task_type = eval_item['tags'][0] # primary task type
-        chat_history = eval_item['history']
-        last_query = eval_item['last_query']
-        # print(f"## Task Type: {task_type}")
-        # print(f"## Chat History: {chat_history}")
-        # print(f"## Last Query -->  USER: {last_query}")
-
-        model_A_output = eval_item['model_1_output'] if model_1 == model_A else eval_item['model_2_output']
-        model_B_output = eval_item['model_2_output'] if model_2 == model_B else eval_item['model_1_output']
-
-        if len(model_A_output.strip()) == 0 or len(model_B_output.strip()) == 0:
-            continue
-
-        conversation_input = id_to_data[eval_item['session_id']]["conversation_input"]
-        # print(f"\n\n\n## Model A ({model_A}) Output ##\n{model_A_output}")
-        # print(f"\n\n\n## Model B ({model_B}) Output ##\n{model_B_output}")
-
-        # print(f"\n\n\n## Winner ##\n{winner}")
-        # print(f"\n\n\n## GPT-4 Judgement ##\n{eval_item['parsed_result']}")
-
-        result_dict = {
-            "session_id": eval_item['session_id'],
-            "model_A": model_A,
-            "model_B": model_B,
-            "winner": winner,
-            "intent": id_to_data[eval_item['session_id']]["intent"],
-            "task_type": task_type,
-            "all_tags": eval_item['tags'],
-            "chat_history": chat_history,
-            "last_query": last_query,
-            "conversation_input": conversation_input,
-            "model_A_output": model_A_output,
-            "model_B_output": model_B_output,
-            "reason": eval_item['parsed_result']["reason"],
-            "choice": eval_item['parsed_result']["choice"],
-            "checklist": id_to_data[eval_item['session_id']]["checklist"],
-        }
-        break 
-    return result_dict
-
-#id_to_data = load_benchdata_dict()