Spaces:
Sleeping
Sleeping
Trying to add a 'totalling' option in the scroll down
Browse files
app.py
CHANGED
@@ -10,11 +10,10 @@ from games_registry import GAMES_REGISTRY
|
|
10 |
from simulators.base_simulator import PlayerType
|
11 |
from typing import Dict
|
12 |
|
13 |
-
# Extract available LLM models
|
14 |
llm_models = list(LLM_REGISTRY.keys())
|
15 |
|
16 |
-
#
|
17 |
-
#games_list = list(GAMES_REGISTRY.keys())
|
18 |
games_list = [
|
19 |
"rock_paper_scissors",
|
20 |
"prisoners_dilemma",
|
@@ -24,32 +23,21 @@ games_list = [
|
|
24 |
"kuhn_poker",
|
25 |
]
|
26 |
|
27 |
-
#
|
28 |
-
|
29 |
-
|
30 |
-
def generate_stats_file(model_name: str):
|
31 |
-
"""Generate a JSON file with detailed statistics for the selected LLM model."""
|
32 |
-
file_path = f"{model_name}_stats.json"
|
33 |
-
with open(file_path, "w") as f:
|
34 |
-
json.dump(results_tracker.get(model_name, {}), f, indent=4)
|
35 |
-
return file_path
|
36 |
-
|
37 |
-
def provide_download_file(model_name):
|
38 |
-
"""Creates a downloadable JSON file with stats for the selected model."""
|
39 |
-
return generate_stats_file(model_name)
|
40 |
|
41 |
-
|
42 |
-
|
43 |
-
return calculate_leaderboard(game_dropdown.value)
|
44 |
|
45 |
# Load or initialize the results tracker
|
46 |
if os.path.exists(RESULTS_TRACKER_FILE):
|
47 |
with open(RESULTS_TRACKER_FILE, "r") as f:
|
48 |
results_tracker = json.load(f)
|
49 |
else:
|
|
|
50 |
results_tracker = {
|
51 |
llm: {game: {"games": 0, "moves/game": 0, "illegal-moves": 0,
|
52 |
-
"win-rate": 0, "vs Random": 0} for game in games_list}
|
53 |
for llm in llm_models
|
54 |
}
|
55 |
|
@@ -58,28 +46,64 @@ def save_results_tracker():
|
|
58 |
with open(RESULTS_TRACKER_FILE, "w") as f:
|
59 |
json.dump(results_tracker, f, indent=4)
|
60 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
def calculate_leaderboard(selected_game: str) -> pd.DataFrame:
|
62 |
-
"""
|
63 |
-
|
64 |
-
|
65 |
-
|
|
|
|
|
|
|
|
|
|
|
66 |
|
67 |
for llm in llm_models:
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
76 |
|
77 |
leaderboard_df = leaderboard_df.reset_index()
|
78 |
leaderboard_df.rename(columns={"index": "LLM Model"}, inplace=True)
|
79 |
return leaderboard_df
|
80 |
|
81 |
def play_game(game_name, player1_type, player2_type, player1_model, player2_model, rounds):
|
82 |
-
"""
|
83 |
llms = {}
|
84 |
if player1_type == "llm":
|
85 |
llms["Player 1"] = player1_model
|
@@ -91,7 +115,7 @@ def play_game(game_name, player1_type, player2_type, player1_model, player2_mode
|
|
91 |
game_states = []
|
92 |
|
93 |
def log_fn(state):
|
94 |
-
"""
|
95 |
current_player = state.current_player()
|
96 |
legal_moves = state.legal_actions(current_player)
|
97 |
board = str(state)
|
@@ -102,10 +126,11 @@ def play_game(game_name, player1_type, player2_type, player1_model, player2_mode
|
|
102 |
|
103 |
# Gradio Interface
|
104 |
with gr.Blocks() as interface:
|
|
|
105 |
with gr.Tab("Game Arena"):
|
106 |
gr.Markdown("# LLM Game Arena\nSelect a game and players to play against LLMs.")
|
107 |
|
108 |
-
game_dropdown = gr.Dropdown(choices=games_list, label="Select a Game", value=games_list[
|
109 |
player1_dropdown = gr.Dropdown(choices=["human", "random_bot", "llm"], label="Player 1 Type", value="llm")
|
110 |
player2_dropdown = gr.Dropdown(choices=["human", "random_bot", "llm"], label="Player 2 Type", value="random_bot")
|
111 |
player1_model_dropdown = gr.Dropdown(choices=llm_models, label="Player 1 Model", visible=False)
|
@@ -120,17 +145,18 @@ with gr.Blocks() as interface:
|
|
120 |
outputs=result_output,
|
121 |
)
|
122 |
|
|
|
123 |
with gr.Tab("Leaderboard"):
|
124 |
gr.Markdown("# LLM Model Leaderboard\nTrack performance across different games!")
|
125 |
|
126 |
-
game_dropdown = gr.Dropdown(choices=games_list, label="Select Game", value=
|
127 |
-
leaderboard_table = gr.Dataframe(value=calculate_leaderboard(
|
128 |
model_dropdown = gr.Dropdown(choices=llm_models, label="Select LLM Model")
|
129 |
download_button = gr.File(label="Download Statistics File")
|
130 |
refresh_button = gr.Button("Refresh Leaderboard")
|
131 |
|
132 |
def update_leaderboard(selected_game):
|
133 |
-
"""Updates the leaderboard
|
134 |
return calculate_leaderboard(selected_game)
|
135 |
|
136 |
model_dropdown.change(fn=provide_download_file, inputs=[model_dropdown], outputs=[download_button])
|
|
|
10 |
from simulators.base_simulator import PlayerType
|
11 |
from typing import Dict
|
12 |
|
13 |
+
# Extract available LLM models from the registry
|
14 |
llm_models = list(LLM_REGISTRY.keys())
|
15 |
|
16 |
+
# List of available games (manually defined for now)
|
|
|
17 |
games_list = [
|
18 |
"rock_paper_scissors",
|
19 |
"prisoners_dilemma",
|
|
|
23 |
"kuhn_poker",
|
24 |
]
|
25 |
|
26 |
+
# Special leaderboard option for aggregating stats across all games
|
27 |
+
games_list.insert(0, "Total Performance")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
|
29 |
+
# File to persist game results
|
30 |
+
RESULTS_TRACKER_FILE = "results_tracker.json"
|
|
|
31 |
|
32 |
# Load or initialize the results tracker
|
33 |
if os.path.exists(RESULTS_TRACKER_FILE):
|
34 |
with open(RESULTS_TRACKER_FILE, "r") as f:
|
35 |
results_tracker = json.load(f)
|
36 |
else:
|
37 |
+
# Initialize tracking for all LLMs and games
|
38 |
results_tracker = {
|
39 |
llm: {game: {"games": 0, "moves/game": 0, "illegal-moves": 0,
|
40 |
+
"win-rate": 0, "vs Random": 0} for game in games_list[1:]}
|
41 |
for llm in llm_models
|
42 |
}
|
43 |
|
|
|
46 |
with open(RESULTS_TRACKER_FILE, "w") as f:
|
47 |
json.dump(results_tracker, f, indent=4)
|
48 |
|
49 |
+
def generate_stats_file(model_name: str) -> str:
|
50 |
+
"""Generate a JSON file with detailed statistics for the selected LLM model."""
|
51 |
+
file_path = f"{model_name}_stats.json"
|
52 |
+
with open(file_path, "w") as f:
|
53 |
+
json.dump(results_tracker.get(model_name, {}), f, indent=4)
|
54 |
+
return file_path
|
55 |
+
|
56 |
+
def provide_download_file(model_name):
|
57 |
+
"""Creates a downloadable JSON file with stats for the selected model."""
|
58 |
+
return generate_stats_file(model_name)
|
59 |
+
|
60 |
def calculate_leaderboard(selected_game: str) -> pd.DataFrame:
|
61 |
+
"""
|
62 |
+
Generate a structured leaderboard table.
|
63 |
+
- If a specific game is selected, returns performance stats per LLM for that game.
|
64 |
+
- If 'Total Performance' is selected, aggregates stats across all games.
|
65 |
+
"""
|
66 |
+
leaderboard_df = pd.DataFrame(
|
67 |
+
index=llm_models,
|
68 |
+
columns=["# games", "moves/game", "illegal-moves", "win-rate", "vs Random"]
|
69 |
+
)
|
70 |
|
71 |
for llm in llm_models:
|
72 |
+
if selected_game == "Total Performance":
|
73 |
+
# Aggregate stats across all games
|
74 |
+
total_games = sum(results_tracker[llm][game]["games"] for game in games_list[1:])
|
75 |
+
total_moves = sum(results_tracker[llm][game]["moves/game"] * results_tracker[llm][game]["games"]
|
76 |
+
for game in games_list[1:])
|
77 |
+
total_illegal_moves = sum(results_tracker[llm][game]["illegal-moves"] for game in games_list[1:])
|
78 |
+
avg_win_rate = sum(results_tracker[llm][game]["win-rate"] * results_tracker[llm][game]["games"]
|
79 |
+
for game in games_list[1:]) / total_games if total_games > 0 else 0
|
80 |
+
avg_vs_random = sum(results_tracker[llm][game]["vs Random"] * results_tracker[llm][game]["games"]
|
81 |
+
for game in games_list[1:]) / total_games if total_games > 0 else 0
|
82 |
+
|
83 |
+
leaderboard_df.loc[llm] = [
|
84 |
+
total_games,
|
85 |
+
f"{(total_moves / total_games) if total_games > 0 else 0:.1f}",
|
86 |
+
total_illegal_moves,
|
87 |
+
f"{avg_win_rate:.1f}%",
|
88 |
+
f"{avg_vs_random:.1f}%"
|
89 |
+
]
|
90 |
+
else:
|
91 |
+
# Retrieve stats for the selected game
|
92 |
+
game_stats = results_tracker[llm].get(selected_game, {})
|
93 |
+
leaderboard_df.loc[llm] = [
|
94 |
+
game_stats.get("games", 0),
|
95 |
+
game_stats.get("moves/game", 0),
|
96 |
+
game_stats.get("illegal-moves", 0),
|
97 |
+
f"{game_stats.get('win-rate', 0):.1f}%",
|
98 |
+
f"{game_stats.get('vs Random', 0):.1f}%"
|
99 |
+
]
|
100 |
|
101 |
leaderboard_df = leaderboard_df.reset_index()
|
102 |
leaderboard_df.rename(columns={"index": "LLM Model"}, inplace=True)
|
103 |
return leaderboard_df
|
104 |
|
105 |
def play_game(game_name, player1_type, player2_type, player1_model, player2_model, rounds):
|
106 |
+
"""Simulates a game session with the chosen players and logs results."""
|
107 |
llms = {}
|
108 |
if player1_type == "llm":
|
109 |
llms["Player 1"] = player1_model
|
|
|
115 |
game_states = []
|
116 |
|
117 |
def log_fn(state):
|
118 |
+
"""Logs the current game state and available moves."""
|
119 |
current_player = state.current_player()
|
120 |
legal_moves = state.legal_actions(current_player)
|
121 |
board = str(state)
|
|
|
126 |
|
127 |
# Gradio Interface
|
128 |
with gr.Blocks() as interface:
|
129 |
+
# Game Arena Tab
|
130 |
with gr.Tab("Game Arena"):
|
131 |
gr.Markdown("# LLM Game Arena\nSelect a game and players to play against LLMs.")
|
132 |
|
133 |
+
game_dropdown = gr.Dropdown(choices=games_list[1:], label="Select a Game", value=games_list[1])
|
134 |
player1_dropdown = gr.Dropdown(choices=["human", "random_bot", "llm"], label="Player 1 Type", value="llm")
|
135 |
player2_dropdown = gr.Dropdown(choices=["human", "random_bot", "llm"], label="Player 2 Type", value="random_bot")
|
136 |
player1_model_dropdown = gr.Dropdown(choices=llm_models, label="Player 1 Model", visible=False)
|
|
|
145 |
outputs=result_output,
|
146 |
)
|
147 |
|
148 |
+
# Leaderboard Tab
|
149 |
with gr.Tab("Leaderboard"):
|
150 |
gr.Markdown("# LLM Model Leaderboard\nTrack performance across different games!")
|
151 |
|
152 |
+
game_dropdown = gr.Dropdown(choices=games_list, label="Select Game", value="Total Performance")
|
153 |
+
leaderboard_table = gr.Dataframe(value=calculate_leaderboard("Total Performance"), label="Leaderboard")
|
154 |
model_dropdown = gr.Dropdown(choices=llm_models, label="Select LLM Model")
|
155 |
download_button = gr.File(label="Download Statistics File")
|
156 |
refresh_button = gr.Button("Refresh Leaderboard")
|
157 |
|
158 |
def update_leaderboard(selected_game):
|
159 |
+
"""Updates the leaderboard based on the selected game."""
|
160 |
return calculate_leaderboard(selected_game)
|
161 |
|
162 |
model_dropdown.change(fn=provide_download_file, inputs=[model_dropdown], outputs=[download_button])
|