reviewer-arena / aws_utils.py
openreviewer's picture
Upload folder using huggingface_hub
38a86d9 verified
import boto3
import uuid
import datetime
import os
from decimal import Decimal, getcontext
from dotenv import load_dotenv
try:
load_dotenv()
except:
pass
# Load AWS credentials from environment variables
aws_access_key_id = os.environ.get('AWS_ACCESS_KEY_ID')
aws_secret_access_key = os.environ.get('AWS_SECRET_ACCESS_KEY')
aws_region = os.environ.get('AWS_REGION')
# Initialize the DynamoDB client
dynamodb = boto3.resource('dynamodb',
region_name=aws_region,
aws_access_key_id=aws_access_key_id,
aws_secret_access_key=aws_secret_access_key)
# Define the tables
requests_table = dynamodb.Table('reviewer_arena_requests')
leaderboards_table = dynamodb.Table('reviewer_arena_leaderboard')
# Function to write a request to the Requests table
def write_request(user_id, paper_id, model_a, model_b, vote):
request_id = str(uuid.uuid4())
timestamp = str(Decimal(datetime.datetime.now().timestamp()))
response = requests_table.put_item(
Item={
'RequestID': request_id,
'Timestamp': timestamp,
'UserID': user_id,
'PaperID': paper_id,
'ModelA': model_a,
'ModelB': model_b,
'Vote': vote
}
)
return response
# Function to update leaderboard after a vote
def update_leaderboard(model_a, model_b, vote):
# Map vote options to simpler keys
vote_mapping = {
"πŸ‘ A is better": "A is better",
"πŸ‘ B is better": "B is better",
"πŸ‘” Tie": "Tie",
"πŸ‘Ž Both are bad": "Tie" # Assuming "Both are bad" is treated as a tie
}
vote = vote_mapping.get(vote, "Tie") # Default to "Tie" if vote is not found
# Retrieve current stats for ModelA and ModelB
model_a_stats = leaderboards_table.get_item(Key={'ModelID': model_a}).get('Item', {})
model_b_stats = leaderboards_table.get_item(Key={'ModelID': model_b}).get('Item', {})
# Initialize stats if they don't exist
if not model_a_stats:
model_a_stats = {'ModelID': model_a, 'Wins': 0, 'Losses': 0, 'Ties': 0, 'EloScore': Decimal(1200), 'Votes': 0}
leaderboards_table.put_item(Item=model_a_stats)
if not model_b_stats:
model_b_stats = {'ModelID': model_b, 'Wins': 0, 'Losses': 0, 'Ties': 0, 'EloScore': Decimal(1200), 'Votes': 0}
leaderboards_table.put_item(Item=model_b_stats)
# Update stats based on the vote
update_expressions = {
"A is better": {
"model_a": "SET Wins = Wins + :inc, Votes = Votes + :inc",
"model_b": "SET Losses = Losses + :inc, Votes = Votes + :inc"
},
"B is better": {
"model_a": "SET Losses = Losses + :inc, Votes = Votes + :inc",
"model_b": "SET Wins = Wins + :inc, Votes = Votes + :inc"
},
"Tie": {
"model_a": "SET Ties = Ties + :inc, Votes = Votes + :inc",
"model_b": "SET Ties = Ties + :inc, Votes = Votes + :inc"
}
}
expression_a = update_expressions[vote]["model_a"]
expression_b = update_expressions[vote]["model_b"]
# Update ModelA stats
leaderboards_table.update_item(
Key={'ModelID': model_a},
UpdateExpression=expression_a,
ExpressionAttributeValues={':inc': 1}
)
# Update ModelB stats
leaderboards_table.update_item(
Key={'ModelID': model_b},
UpdateExpression=expression_b,
ExpressionAttributeValues={':inc': 1}
)
# Calculate new Elo scores (simple Elo calculation for illustration)
new_elo_a, new_elo_b = calculate_elo(model_a_stats['EloScore'], model_b_stats['EloScore'], vote)
# Calculate 95% CI for new Elo scores
ci_a_lower, ci_a_upper = calculate_95_ci(new_elo_a, model_a_stats['Votes'] + 1)
ci_b_lower, ci_b_upper = calculate_95_ci(new_elo_b, model_b_stats['Votes'] + 1)
# Update Elo scores and 95% CI
leaderboards_table.update_item(
Key={'ModelID': model_a},
UpdateExpression="SET EloScore = :new_elo, CI_Lower = :ci_lower, CI_Upper = :ci_upper",
ExpressionAttributeValues={':new_elo': Decimal(new_elo_a), ':ci_lower': Decimal(ci_a_lower), ':ci_upper': Decimal(ci_a_upper)}
)
leaderboards_table.update_item(
Key={'ModelID': model_b},
UpdateExpression="SET EloScore = :new_elo, CI_Lower = :ci_lower, CI_Upper = :ci_upper",
ExpressionAttributeValues={':new_elo': Decimal(new_elo_b), ':ci_lower': Decimal(ci_b_lower), ':ci_upper': Decimal(ci_b_upper)}
)
# Set the precision for Decimal
getcontext().prec = 28
# Function to calculate new Elo scores
def calculate_elo(elo_a, elo_b, vote, k=32):
# Ensure elo_a and elo_b are Decimals
elo_a = Decimal(elo_a)
elo_b = Decimal(elo_b)
expected_a = 1 / (1 + Decimal(10) ** ((elo_b - elo_a) / Decimal(400)))
expected_b = 1 / (1 + Decimal(10) ** ((elo_a - elo_b) / Decimal(400)))
if vote == "A is better":
actual_a = Decimal(1)
actual_b = Decimal(0)
elif vote == "B is better":
actual_a = Decimal(0)
actual_b = Decimal(1)
else: # Tie
actual_a = Decimal(0.5)
actual_b = Decimal(0.5)
new_elo_a = elo_a + Decimal(k) * (actual_a - expected_a)
new_elo_b = elo_b + Decimal(k) * (actual_b - expected_b)
return round(new_elo_a, 2), round(new_elo_b, 2)
# Function to calculate 95% CI for Elo scores
def calculate_95_ci(elo, votes, z=1.96):
if votes == 0:
return Decimal(0), Decimal(0)
elo = Decimal(elo) # Ensure elo is a Decimal
std_error = Decimal(400) / (Decimal(votes).sqrt())
margin = Decimal(z) * std_error
return round(elo - margin, 2), round(elo + margin, 2)
# Function to query leaderboard
def get_leaderboard():
response = leaderboards_table.scan()
leaderboard = response.get('Items', [])
# Sort by EloScore in descending order
leaderboard.sort(key=lambda x: x['EloScore'], reverse=True)
return leaderboard