MMLU-by-task-Leaderboard / result_data_processor.py
Corey Morris
Returning just a single file per model directory. Manually removing gpt-j-6b for now because there is something that is causing problems with processing the data
794b32b
raw
history blame
6.19 kB
import pandas as pd
import os
import fnmatch
import json
import re
import numpy as np
class ResultDataProcessor:
def __init__(self, directory='results', pattern='results*.json'):
self.directory = directory
self.pattern = pattern
self.data = self.process_data()
self.ranked_data = self.rank_data()
def _find_files(self, directory='results', pattern='results*.json'):
matching_files = {}
for root, dirs, files in os.walk(directory):
for basename in files:
if fnmatch.fnmatch(basename, pattern):
filename = os.path.join(root, basename)
matching_files[root] = filename
matching_files = {key: value for key, value in matching_files.items() if 'gpt-j-6b' not in key}
matching_files = list(matching_files.values())
return matching_files
def _read_and_transform_data(self, filename):
with open(filename) as f:
data = json.load(f)
df = pd.DataFrame(data['results']).T
return df
def _cleanup_dataframe(self, df, model_name):
df = df.rename(columns={'acc': model_name})
df.index = (df.index.str.replace('hendrycksTest-', 'MMLU_', regex=True)
.str.replace('harness\|', '', regex=True)
.str.replace('\|5', '', regex=True))
return df[[model_name]]
def _extract_mc1(self, df, model_name):
df = df.rename(columns={'mc1': model_name})
# rename row harness|truthfulqa:mc|0 to truthfulqa:mc1
df.index = (df.index.str.replace('mc\|0', 'mc1', regex=True))
# just return the harness|truthfulqa:mc1 row
df = df.loc[['harness|truthfulqa:mc1']]
return df[[model_name]]
def _extract_mc2(self, df, model_name):
# rename row harness|truthfulqa:mc|0 to truthfulqa:mc2
df = df.rename(columns={'mc2': model_name})
df.index = (df.index.str.replace('mc\|0', 'mc2', regex=True))
df = df.loc[['harness|truthfulqa:mc2']]
return df[[model_name]]
# remove extreme outliers from column harness|truthfulqa:mc1
def _remove_mc1_outliers(self, df):
mc1 = df['harness|truthfulqa:mc1']
# Identify the outliers
# outliers_condition = mc1 > mc1.quantile(.95)
outliers_condition = mc1 == 1.0
# Replace the outliers with NaN
df.loc[outliers_condition, 'harness|truthfulqa:mc1'] = np.nan
return df
@staticmethod
def _extract_parameters(model_name):
"""
Function to extract parameters from model name.
It handles names with 'b/B' for billions and 'm/M' for millions.
"""
# pattern to match a number followed by 'b' (representing billions) or 'm' (representing millions)
pattern = re.compile(r'(\d+\.?\d*)([bBmM])')
match = pattern.search(model_name)
if match:
num, magnitude = match.groups()
num = float(num)
# convert millions to billions
if magnitude.lower() == 'm':
num /= 1000
return num
# return NaN if no match
return np.nan
def process_data(self):
dataframes = []
organization_names = []
for filename in self._find_files(self.directory, self.pattern):
raw_data = self._read_and_transform_data(filename)
split_path = filename.split('/')
model_name = split_path[2]
organization_name = split_path[1]
cleaned_data = self._cleanup_dataframe(raw_data, model_name)
mc1 = self._extract_mc1(raw_data, model_name)
mc2 = self._extract_mc2(raw_data, model_name)
cleaned_data = pd.concat([cleaned_data, mc1])
cleaned_data = pd.concat([cleaned_data, mc2])
organization_names.append(organization_name)
dataframes.append(cleaned_data)
data = pd.concat(dataframes, axis=1).transpose()
# Add organization column
data['organization'] = organization_names
# Add Model Name and rearrange columns
data['Model Name'] = data.index
cols = data.columns.tolist()
cols = cols[-1:] + cols[:-1]
data = data[cols]
# Remove the 'Model Name' column
data = data.drop(columns=['Model Name'])
# Add average column
data['MMLU_average'] = data.filter(regex='MMLU').mean(axis=1)
# Reorder columns to move 'MMLU_average' to the third position
cols = data.columns.tolist()
cols = cols[:2] + cols[-1:] + cols[2:-1]
data = data[cols]
# Drop specific columns
data = data.drop(columns=['all', 'truthfulqa:mc|0'])
# Add parameter count column using extract_parameters function
data['Parameters'] = data.index.to_series().apply(self._extract_parameters)
# move the parameters column to the front of the dataframe
cols = data.columns.tolist()
cols = cols[-1:] + cols[:-1]
data = data[cols]
# remove extreme outliers from column harness|truthfulqa:mc1
data = self._remove_mc1_outliers(data)
data = self.manual_removal_of_models(data)
return data
def manual_removal_of_models(self, df):
# remove models verified to be trained on evaluation data
# load the list of models
with open('contaminated_models.txt') as f:
contaminated_models = f.read().splitlines()
# remove the models from the dataframe
df = df[~df.index.isin(contaminated_models)]
return df
def rank_data(self):
# add rank for each column to the dataframe
# copy the data dataframe to avoid modifying the original dataframe
rank_data = self.data.copy()
for col in list(rank_data.columns):
rank_data[col + "_rank"] = rank_data[col].rank(ascending=False, method='min')
return rank_data
def get_data(self, selected_models):
return self.data[self.data.index.isin(selected_models)]