Spaces:

GTBench
/

GTBench

Running

GTBench / app.py

Renming Zhang

Merge commit 'bc5a5291ead7e73ca497b6d570e64c808b871bb9'

9c2a184 3 months ago

No virus

18.1 kB

	import gradio as gr
	import pandas as pd
	from src.display.about import (
	CITATION_BUTTON_LABEL,
	CITATION_BUTTON_TEXT,
	EVALUATION_QUEUE_TEXT,
	INTRODUCTION_TEXT,
	LLM_BENCHMARKS_TEXT,
	FAQ_TEXT,
	TITLE,
	)
	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, EVAL_REQUESTS_PATH, EVAL_RESULTS_PATH, H4_TOKEN, IS_PUBLIC, QUEUE_REPO, REPO_ID, RESULTS_REPO
	from PIL import Image
	# from src.populate import get_evaluation_queue_df, get_leaderboard_df
	# from src.submission.submit import add_new_eval
	# from src.tools.collections import update_collections
	# from src.tools.plots import (
	# create_metric_plot_obj,
	# create_plot_df,
	# create_scores_df,
	# )
	from dummydatagen import dummy_data_for_plot, create_metric_plot_obj_1, dummydf
	import copy


	def restart_space():
	API.restart_space(repo_id=REPO_ID, token=H4_TOKEN)


	def add_average_col(df):

	always_here_cols = [
	"Model", "Agent", "Opponent Model", "Opponent Agent"
	]
	desired_col = [i for i in list(df.columns) if i not in always_here_cols]
	newdf = df[desired_col].mean(axis=1).round(3)
	return newdf


	gtbench_raw_data = dummydf()
	gtbench_raw_data["Average"] = add_average_col(gtbench_raw_data)

	column_to_move = "Average"
	# Move the column to the desired index
	gtbench_raw_data.insert(
	4, column_to_move, gtbench_raw_data.pop(column_to_move))

	models = list(set(gtbench_raw_data['Model']))

	opponent_models = list(set(gtbench_raw_data['Opponent Model']))


	agents = list(set(gtbench_raw_data['Agent']))


	opponent_agents = list(set(gtbench_raw_data['Opponent Agent']))

	# Searching and filtering


	def update_table(
	hidden_df: pd.DataFrame,
	columns: list,
	model1: list,
	model2: list,
	agent1: list,
	agent2: list
	):

	filtered_df = select_columns(hidden_df, columns)

	filtered_df = filter_model1(filtered_df, model1)
	filtered_df = filter_model2(filtered_df, model2)
	filtered_df = filter_agent1(filtered_df, agent1)
	filtered_df = filter_agent2(filtered_df, agent2)

	return filtered_df

	# triggered only once at startup => read query parameter if it exists


	def load_query(request: gr.Request):
	query = request.query_params.get("query") or ""
	return query, query # return one for the "search_bar", one for a hidden component that triggers a reload only if value has changed


	def search_table(df: pd.DataFrame, query: str) -> pd.DataFrame:
	return df[(df[AutoEvalColumn.dummy.name].str.contains(query, case=False))]


	def select_columns(df: pd.DataFrame, columns: list) -> pd.DataFrame:
	always_here_cols = [
	"Model", "Agent", "Opponent Model", "Opponent Agent"
	]
	# We use COLS to maintain sorting
	all_columns = games

	if len(columns) == 0:
	filtered_df = df[
	always_here_cols +
	[c for c in all_columns if c in df.columns]
	]
	filtered_df["Average"] = add_average_col(filtered_df)
	column_to_move = "Average"
	current_index = filtered_df.columns.get_loc(column_to_move)

	# Move the column to the desired index
	filtered_df.insert(4, column_to_move, filtered_df.pop(column_to_move))
	return filtered_df

	filtered_df = df[
	always_here_cols +
	[c for c in all_columns if c in df.columns and c in columns]
	]
	if "Average" in columns:
	filtered_df["Average"] = add_average_col(filtered_df)
	# Get the current index of the column
	column_to_move = "Average"
	current_index = filtered_df.columns.get_loc(column_to_move)

	# Move the column to the desired index
	filtered_df.insert(4, column_to_move, filtered_df.pop(column_to_move))
	else:
	if "Average" in filtered_df.columns:
	# Remove the column
	filtered_df = filtered_df.drop(columns=["Average"])

	return filtered_df


	def filter_model1(
	df: pd.DataFrame, model_query: list
	) -> pd.DataFrame:
	# Show all models
	if len(model_query) == 0:
	return df
	filtered_df = df

	filtered_df = filtered_df[filtered_df["Model"].isin(
	model_query)]
	return filtered_df


	def filter_model2(
	df: pd.DataFrame, model_query: list
	) -> pd.DataFrame:
	# Show all models
	if len(model_query) == 0:
	return df
	filtered_df = df

	filtered_df = filtered_df[filtered_df["Opponent Model"].isin(
	model_query)]
	return filtered_df


	def filter_agent1(
	df: pd.DataFrame, agent_query: list
	) -> pd.DataFrame:
	# Show all models
	if len(agent_query) == 0:
	return df
	filtered_df = df

	filtered_df = filtered_df[filtered_df["Agent"].isin(
	agent_query)]
	return filtered_df


	def filter_agent2(
	df: pd.DataFrame, agent_query: list
	) -> pd.DataFrame:
	# Show all models
	if len(agent_query) == 0:
	return df
	filtered_df = df

	filtered_df = filtered_df[filtered_df["Opponent Agent"].isin(
	agent_query)]
	return filtered_df


	# leaderboard_df = filter_models(leaderboard_df, [t.to_str(" : ") for t in ModelType], list(NUMERIC_INTERVALS.keys()), [i.value.name for i in Precision], False, False)


	class LLM_Model:
	def __init__(self, t_value, model_value, average_value, arc_value, hellaSwag_value, mmlu_value) -> None:
	self.t = t_value
	self.model = model_value
	self.average = average_value
	self.arc = arc_value
	self.hellaSwag = hellaSwag_value
	self.mmlu = mmlu_value


	games = ["Breakthrough", "Connect Four", "Blind Auction", "Kuhn Poker",
	"Liar's Dice", "Negotiation", "Nim", "Pig", "Iterated Prisoner's Dilemma", "Tic-Tac-Toe"]

	# models = ["gpt-35-turbo-1106", "gpt-4", "Llama-2-70b-chat-hf", "CodeLlama-34b-Instruct-hf",
	# "CodeLlama-70b-Instruct-hf", "Mistral-7B-Instruct-v01", "Mistral-7B-OpenOrca"]

	# agents = ["Prompt Agent", "CoT Agent", "SC-CoT Agent",
	# "ToT Agent", "MCTS", "Random", "TitforTat"]

	demo = gr.Blocks(css=custom_css)


	def load_image(image_path):
	image = Image.open(image_path)
	return image


	with demo:
	with gr.Row():
	gr.Image("./assets/logo.png", height="200px", width="200px", scale=0.1,
	show_download_button=False, container=False)
	gr.HTML(TITLE, elem_id="title")

	gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
	with gr.Tabs(elem_classes="tab-buttons") as tabs:
	with gr.TabItem("🏅 GTBench", elem_id="llm-benchmark-tab-table", id=0):
	with gr.Row():
	with gr.Column():
	with gr.Row():

	shown_columns = gr.CheckboxGroup(
	choices=[
	'Average'
	]+games,
	label="Select columns to show",
	elem_id="column-select",
	interactive=True,
	)
	with gr.Column(min_width=320):
	# with gr.Box(elem_id="box-filter"):
	model1_column = gr.CheckboxGroup(
	label="Model",
	choices=models,
	interactive=True,
	elem_id="filter-columns-type",
	)

	agent1_column = gr.CheckboxGroup(
	label="Agents",
	choices=agents,
	interactive=True,
	elem_id="filter-columns-precision",
	)

	model2_column = gr.CheckboxGroup(
	label="Opponent Model",
	choices=opponent_models,
	interactive=True,
	elem_id="filter-columns-type",
	)
	agent2_column = gr.CheckboxGroup(
	label="Opponent Agents",
	choices=opponent_agents,
	interactive=True,
	elem_id="filter-columns-precision",
	)
	# filter_columns_size = gr.CheckboxGroup(
	# label="Model sizes (in billions of parameters)",
	# choices=[f'NUMERIC_INTERVALS{i}' for i in range(0, 5)],
	# value=[f'NUMERIC_INTERVALS{i}' for i in range(0, 5)],
	# interactive=True,
	# elem_id="filter-columns-size",
	# )

	leaderboard_table = gr.components.Dataframe(
	value=gtbench_raw_data,
	elem_id="leaderboard-table",
	interactive=False,
	visible=True,
	# column_widths=["2%", "33%"]
	)

	game_bench_df_for_search = gr.components.Dataframe(
	value=gtbench_raw_data,
	elem_id="leaderboard-table",
	interactive=False,
	visible=False,
	# column_widths=["2%", "33%"]
	)

	# Dummy leaderboard for handling the case when the user uses backspace key
	# hidden_leaderboard_table_for_search = gr.components.Dataframe(
	# value=[],
	# headers=COLS,
	# datatype=TYPES,
	# visible=False,
	# )
	# search_bar.submit(
	# update_table,
	# [
	# # hidden_leaderboard_table_for_search,
	# # shown_columns,
	# # filter_columns_type,
	# # filter_columns_precision,
	# # filter_columns_size,
	# # deleted_models_visibility,
	# # flagged_models_visibility,
	# # search_bar,
	# ],
	# leaderboard_table,
	# )

	# # Define a hidden component that will trigger a reload only if a query parameter has be set
	# hidden_search_bar = gr.Textbox(value="", visible=False)
	# hidden_search_bar.change(
	# update_table,
	# [
	# hidden_leaderboard_table_for_search,
	# shown_columns,
	# filter_columns_type,
	# filter_columns_precision,
	# filter_columns_size,
	# deleted_models_visibility,
	# flagged_models_visibility,
	# search_bar,
	# ],
	# leaderboard_table,
	# )
	# # Check query parameter once at startup and update search bar + hidden component
	# demo.load(load_query, inputs=[], outputs=[search_bar, hidden_search_bar])

	for selector in [shown_columns, model1_column, model2_column, agent1_column, agent2_column]:
	selector.change(
	update_table,
	[
	game_bench_df_for_search,
	shown_columns,
	model1_column,
	model2_column,
	agent1_column,
	agent2_column
	# filter_columns_precision,
	# None, # filter_columns_size,
	# None, # deleted_models_visibility,
	# None, # flagged_models_visibility,
	# None, # search_bar,
	],
	leaderboard_table,
	queue=True,
	)

	# with gr.TabItem("📈 Metrics through time", elem_id="llm-benchmark-tab-table", id=4):
	# with gr.Row():
	# with gr.Column():
	# chart = create_metric_plot_obj_1(
	# dummy_data_for_plot(
	# ["Metric1", "Metric2", 'Metric3']),
	# ["Metric1", "Metric2", "Metric3"],
	# title="Average of Top Scores and Human Baseline Over Time (from last update)",
	# )
	# gr.Plot(value=chart, min_width=500)

	with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=2):
	gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
	gr.Markdown(FAQ_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 ({9})",
	open=False,
	):
	with gr.Row():
	finished_eval_table = gr.components.Dataframe(
	value=None,
	headers=EVAL_COLS,
	datatype=EVAL_TYPES,
	row_count=5,
	)
	with gr.Accordion(
	f"🔄 Running Evaluation Queue ({5})",
	open=False,
	):
	with gr.Row():
	running_eval_table = gr.components.Dataframe(
	value=None,
	headers=EVAL_COLS,
	datatype=EVAL_TYPES,
	row_count=5,
	)

	with gr.Accordion(
	f"⏳ Pending Evaluation Queue ({7})",
	open=False,
	):
	with gr.Row():
	pending_eval_table = gr.components.Dataframe(
	value=None,
	headers=EVAL_COLS,
	datatype=EVAL_TYPES,
	row_count=5,
	)
	with gr.Row():
	gr.Markdown("# ✉️✨ Submit your Agent here!",
	elem_classes="markdown-text")

	with gr.Row():
	with gr.Column():
	model_name_textbox = gr.Textbox(label="Agent name")
	# revision_name_textbox = gr.Textbox(
	# label="Revision commit", placeholder="main")
	# private = gr.Checkbox(
	# False, label="Private", visible=not IS_PUBLIC)
	model_type = gr.Dropdown(
	choices=[t.to_str(" : ")
	for t in ModelType if t != ModelType.Unknown],
	label="Agent type",
	multiselect=False,
	value=ModelType.FT.to_str(" : "),
	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,
	# private,
	# 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.launch()
	# Both launches the space and its CI
	# configure_space_ci(
	# demo.queue(default_concurrency_limit=40),
	# trusted_authors=[], # add manually trusted authors
	# private="True", # ephemeral spaces will have same visibility as the main space. Otherwise, set to `True` or `False` explicitly.
	# variables={}, # We overwrite HF_HOME as tmp CI spaces will have no cache
	# secrets=["HF_TOKEN", "H4_TOKEN"], # which secret do I want to copy from the main space? Can be a `List[str]`."HF_TOKEN", "H4_TOKEN"
	# hardware=None, # "cpu-basic" by default. Otherwise set to "auto" to have same hardware as the main space or any valid string value.
	# storage=None, # no storage by default. Otherwise set to "auto" to have same storage as the main space or any valid string value.
	# ).launch()


	# notes: opponent model , opponent agent
	# column is games