File size: 5,519 Bytes
843a5ef
 
 
 
dcadab7
 
843a5ef
 
ee5ac8e
 
 
 
843a5ef
31bed1a
843a5ef
ee5ac8e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
abac22e
 
 
 
 
 
 
 
 
 
 
 
 
 
e03b231
 
 
 
 
 
 
 
 
 
 
abac22e
 
dcadab7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
843a5ef
6d41115
 
e79bcf3
6d41115
 
52d3b03
 
 
6d41115
abac22e
 
 
 
e79bcf3
6d41115
 
843a5ef
ee5ac8e
e79bcf3
 
 
 
ee5ac8e
843a5ef
 
 
 
 
ee5ac8e
 
 
 
843a5ef
 
ee5ac8e
843a5ef
 
 
 
ee5ac8e
9549fcc
dcadab7
 
 
 
 
 
 
 
 
e03b231
 
 
dcadab7
31bed1a
 
 
 
 
 
 
 
 
ee5ac8e
843a5ef
ee5ac8e
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
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()

    @staticmethod
    def _find_files(directory, pattern):
        for root, dirs, files in os.walk(directory):
            for basename in files:
                if fnmatch.fnmatch(basename, pattern):
                    filename = os.path.join(root, basename)
                    yield filename
    
    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)

        return data
    
    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)]