gpu-poor-llm-arena / leaderboard.py
k-mktr's picture
Update leaderboard.py
b0ade41 verified
raw
history blame
15.2 kB
from nc_py_api import Nextcloud
import json
from typing import Dict, Any
import os
import time
from datetime import datetime
import threading
import arena_config
import sys
import math
import plotly.graph_objects as go
# Initialize Nextcloud client
nc = Nextcloud(nextcloud_url=arena_config.NEXTCLOUD_URL, nc_auth_user=arena_config.NEXTCLOUD_USERNAME, nc_auth_pass=arena_config.NEXTCLOUD_PASSWORD)
# Dictionary to store ELO ratings
elo_ratings = {}
def load_leaderboard() -> Dict[str, Any]:
try:
file_content = nc.files.download(arena_config.NEXTCLOUD_LEADERBOARD_PATH)
return json.loads(file_content.decode('utf-8'))
except Exception as e:
print(f"Error loading leaderboard: {str(e)}")
return {}
def save_leaderboard(leaderboard_data: Dict[str, Any]) -> bool:
try:
json_data = json.dumps(leaderboard_data, indent=2)
nc.files.upload(arena_config.NEXTCLOUD_LEADERBOARD_PATH, json_data.encode('utf-8'))
return True
except Exception as e:
print(f"Error saving leaderboard: {str(e)}")
return False
def get_model_size(model_name):
for model, human_readable in arena_config.APPROVED_MODELS:
if model == model_name:
size = float(human_readable.split('(')[1].split('B')[0])
return size
return 1.0 # Default size if not found
def calculate_expected_score(rating_a, rating_b):
return 1 / (1 + math.pow(10, (rating_b - rating_a) / 400))
def update_elo_ratings(winner, loser):
if winner not in elo_ratings or loser not in elo_ratings:
initialize_elo_ratings()
winner_rating = elo_ratings[winner]
loser_rating = elo_ratings[loser]
expected_winner = calculate_expected_score(winner_rating, loser_rating)
expected_loser = 1 - expected_winner
winner_size = get_model_size(winner)
loser_size = get_model_size(loser)
max_size = max(get_model_size(model) for model, _ in arena_config.APPROVED_MODELS)
k_factor = min(64, 32 * (1 + (loser_size - winner_size) / max_size))
elo_ratings[winner] += k_factor * (1 - expected_winner)
elo_ratings[loser] += k_factor * (0 - expected_loser)
def initialize_elo_ratings():
leaderboard = load_leaderboard()
for model, _ in arena_config.APPROVED_MODELS:
size = get_model_size(model)
elo_ratings[model] = 1000 + (size * 100)
# Replay all battles to update ELO ratings
for model, data in leaderboard.items():
if model not in elo_ratings:
elo_ratings[model] = 1000 + (get_model_size(model) * 100)
for opponent, results in data['opponents'].items():
if opponent not in elo_ratings:
elo_ratings[opponent] = 1000 + (get_model_size(opponent) * 100)
for _ in range(results['wins']):
update_elo_ratings(model, opponent)
for _ in range(results['losses']):
update_elo_ratings(opponent, model)
def ensure_elo_ratings_initialized():
if not elo_ratings:
initialize_elo_ratings()
def update_leaderboard(winner: str, loser: str) -> Dict[str, Any]:
leaderboard = load_leaderboard()
if winner not in leaderboard:
leaderboard[winner] = {"wins": 0, "losses": 0, "opponents": {}}
if loser not in leaderboard:
leaderboard[loser] = {"wins": 0, "losses": 0, "opponents": {}}
leaderboard[winner]["wins"] += 1
leaderboard[winner]["opponents"].setdefault(loser, {"wins": 0, "losses": 0})["wins"] += 1
leaderboard[loser]["losses"] += 1
leaderboard[loser]["opponents"].setdefault(winner, {"wins": 0, "losses": 0})["losses"] += 1
# Update ELO ratings
update_elo_ratings(winner, loser)
save_leaderboard(leaderboard)
return leaderboard
def get_current_leaderboard() -> Dict[str, Any]:
return load_leaderboard()
def get_human_readable_name(model_name: str) -> str:
model_dict = dict(arena_config.APPROVED_MODELS)
return model_dict.get(model_name, model_name)
def get_leaderboard():
leaderboard = load_leaderboard()
# Calculate scores for each model
for model, results in leaderboard.items():
total_battles = results["wins"] + results["losses"]
if total_battles > 0:
win_rate = results["wins"] / total_battles
results["score"] = win_rate * (1 - 1 / (total_battles + 1))
else:
results["score"] = 0
# Sort results by score, then by total battles
sorted_results = sorted(
leaderboard.items(),
key=lambda x: (x[1]["score"], x[1]["wins"] + x[1]["losses"]),
reverse=True
)
# Explanation of the main leaderboard
explanation = """
<p style="font-size: 16px; margin-bottom: 20px;">
This leaderboard uses a scoring system that balances win rate and total battles. The score is calculated using the formula:
<br>
<strong>Score = Win Rate * (1 - 1 / (Total Battles + 1))</strong>
<br>
This formula rewards models with higher win rates and more battles. As the number of battles increases, the score approaches the win rate.
</p>
"""
leaderboard_html = f"""
{explanation}
<style>
.leaderboard-table {{
width: 100%;
border-collapse: collapse;
font-family: Arial, sans-serif;
}}
.leaderboard-table th, .leaderboard-table td {{
border: 1px solid #ddd;
padding: 8px;
text-align: left;
}}
.leaderboard-table th {{
background-color: rgba(255, 255, 255, 0.1);
font-weight: bold;
}}
.rank-column {{
width: 60px;
text-align: center;
}}
.opponent-details {{
font-size: 0.9em;
color: #888;
}}
</style>
<table class='leaderboard-table'>
<tr>
<th class='rank-column'>Rank</th>
<th>Model</th>
<th>Score</th>
<th>Wins</th>
<th>Losses</th>
<th>Win Rate</th>
<th>Total Battles</th>
<th>Top Rival</th>
<th>Toughest Opponent</th>
</tr>
"""
for index, (model, results) in enumerate(sorted_results, start=1):
total_battles = results["wins"] + results["losses"]
win_rate = (results["wins"] / total_battles * 100) if total_battles > 0 else 0
rank_display = {1: "πŸ₯‡", 2: "πŸ₯ˆ", 3: "πŸ₯‰"}.get(index, f"{index}")
top_rival = max(results["opponents"].items(), key=lambda x: x[1]["wins"], default=(None, {"wins": 0}))
top_rival_name = get_human_readable_name(top_rival[0]) if top_rival[0] else "N/A"
top_rival_wins = top_rival[1]["wins"]
toughest_opponent = max(results["opponents"].items(), key=lambda x: x[1]["losses"], default=(None, {"losses": 0}))
toughest_opponent_name = get_human_readable_name(toughest_opponent[0]) if toughest_opponent[0] else "N/A"
toughest_opponent_losses = toughest_opponent[1]["losses"]
leaderboard_html += f"""
<tr>
<td class='rank-column'>{rank_display}</td>
<td>{get_human_readable_name(model)}</td>
<td>{results['score']:.4f}</td>
<td>{results['wins']}</td>
<td>{results['losses']}</td>
<td>{win_rate:.2f}%</td>
<td>{total_battles}</td>
<td class='opponent-details'>{top_rival_name} (W: {top_rival_wins})</td>
<td class='opponent-details'>{toughest_opponent_name} (L: {toughest_opponent_losses})</td>
</tr>
"""
leaderboard_html += "</table>"
return leaderboard_html
def calculate_elo_impact(model):
positive_impact = 0
negative_impact = 0
leaderboard = load_leaderboard()
initial_rating = 1000 + (get_model_size(model) * 100)
if model in leaderboard:
for opponent, results in leaderboard[model]['opponents'].items():
model_size = get_model_size(model)
opponent_size = get_model_size(opponent)
max_size = max(get_model_size(m) for m, _ in arena_config.APPROVED_MODELS)
size_difference = (opponent_size - model_size) / max_size
win_impact = 1 + max(0, size_difference)
loss_impact = 1 + max(0, -size_difference)
positive_impact += results['wins'] * win_impact
negative_impact += results['losses'] * loss_impact
return round(positive_impact), round(negative_impact), round(initial_rating)
def get_elo_leaderboard():
ensure_elo_ratings_initialized()
leaderboard = load_leaderboard()
# Create a list of all models, including those from APPROVED_MODELS that might not be in the leaderboard yet
all_models = set(dict(arena_config.APPROVED_MODELS).keys()) | set(leaderboard.keys())
elo_data = []
for model in all_models:
initial_rating = 1000 + (get_model_size(model) * 100)
current_rating = elo_ratings.get(model, initial_rating)
# Calculate battle data only if the model exists in the leaderboard
if model in leaderboard:
wins = leaderboard[model].get('wins', 0)
losses = leaderboard[model].get('losses', 0)
total_battles = wins + losses
positive_impact, negative_impact, _ = calculate_elo_impact(model)
else:
wins = losses = total_battles = positive_impact = negative_impact = 0
elo_data.append({
'model': model,
'current_rating': current_rating,
'initial_rating': initial_rating,
'total_battles': total_battles,
'positive_impact': positive_impact,
'negative_impact': negative_impact
})
# Sort the data by current rating
sorted_elo_data = sorted(elo_data, key=lambda x: x['current_rating'], reverse=True)
min_initial_rating = min(data['initial_rating'] for data in elo_data)
max_initial_rating = max(data['initial_rating'] for data in elo_data)
explanation_elo = f"""
<p style="font-size: 16px; margin-bottom: 20px;">
This leaderboard uses a modified ELO rating system that takes into account both the performance and size of the models.
Initial ratings range from {round(min_initial_rating)} to {round(max_initial_rating)} points, based on model size, with larger models starting at higher ratings.
The "Positive Impact" score reflects the significance of wins, with higher scores for defeating larger models.
The "Negative Impact" score indicates the significance of losses, with higher scores for losing against smaller models.
The current ELO rating is calculated based on these impacts and the model's performance history.
</p>
"""
leaderboard_html = f"""
{explanation_elo}
<style>
.elo-leaderboard-table {{
width: 100%;
border-collapse: collapse;
font-family: Arial, sans-serif;
}}
.elo-leaderboard-table th, .elo-leaderboard-table td {{
border: 1px solid #ddd;
padding: 8px;
text-align: left;
}}
.elo-leaderboard-table th {{
background-color: rgba(255, 255, 255, 0.1);
font-weight: bold;
}}
.rank-column {{
width: 60px;
text-align: center;
}}
</style>
<table class='elo-leaderboard-table'>
<tr>
<th class='rank-column'>Rank</th>
<th>Model</th>
<th>Current ELO Rating</th>
<th>Positive Impact</th>
<th>Negative Impact</th>
<th>Total Battles</th>
<th>Initial Rating</th>
</tr>
"""
for index, data in enumerate(sorted_elo_data, start=1):
rank_display = {1: "πŸ₯‡", 2: "πŸ₯ˆ", 3: "πŸ₯‰"}.get(index, f"{index}")
leaderboard_html += f"""
<tr>
<td class='rank-column'>{rank_display}</td>
<td>{get_human_readable_name(data['model'])}</td>
<td><strong>{round(data['current_rating'])}</strong></td>
<td>{data['positive_impact']}</td>
<td>{data['negative_impact']}</td>
<td>{data['total_battles']}</td>
<td>{round(data['initial_rating'])}</td>
</tr>
"""
leaderboard_html += "</table>"
return leaderboard_html
def create_backup():
while True:
try:
leaderboard_data = load_leaderboard()
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
backup_file_name = f"leaderboard_backup_{timestamp}.json"
backup_path = f"{arena_config.NEXTCLOUD_BACKUP_FOLDER}/{backup_file_name}"
json_data = json.dumps(leaderboard_data, indent=2)
nc.files.upload(backup_path, json_data.encode('utf-8'))
print(f"Backup created on Nextcloud: {backup_path}")
except Exception as e:
print(f"Error creating backup: {e}")
time.sleep(3600) # Sleep for 1 HOUR
def start_backup_thread():
backup_thread = threading.Thread(target=create_backup, daemon=True)
backup_thread.start()
def get_leaderboard_chart():
battle_results = get_current_leaderboard()
# Calculate scores and sort results
for model, results in battle_results.items():
total_battles = results["wins"] + results["losses"]
if total_battles > 0:
win_rate = results["wins"] / total_battles
results["score"] = win_rate * (1 - 1 / (total_battles + 1))
else:
results["score"] = 0
sorted_results = sorted(
battle_results.items(),
key=lambda x: (x[1]["score"], x[1]["wins"] + x[1]["losses"]),
reverse=True
)
models = [get_human_readable_name(model) for model, _ in sorted_results]
wins = [results["wins"] for _, results in sorted_results]
losses = [results["losses"] for _, results in sorted_results]
scores = [results["score"] for _, results in sorted_results]
fig = go.Figure()
# Stacked Bar chart for Wins and Losses
fig.add_trace(go.Bar(
x=models,
y=wins,
name='Wins',
marker_color='#22577a'
))
fig.add_trace(go.Bar(
x=models,
y=losses,
name='Losses',
marker_color='#38a3a5'
))
# Line chart for Scores
fig.add_trace(go.Scatter(
x=models,
y=scores,
name='Score',
yaxis='y2',
line=dict(color='#ff7f0e', width=2)
))
# Update layout for full-width, increased height, and secondary y-axis
fig.update_layout(
title='Model Performance',
xaxis_title='Models',
yaxis_title='Number of Battles',
yaxis2=dict(
title='Score',
overlaying='y',
side='right'
),
barmode='stack',
height=800,
width=1450,
autosize=True,
legend=dict(
orientation='h',
yanchor='bottom',
y=1.02,
xanchor='right',
x=1
)
)
return fig