StrategicDataOrdering/preprocess.py at master · KoJLabs/StrategicDataOrdering · GitHub

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
157
158
159
160
import os
import random

import numpy as np
import pandas as pd

from tqdm import tqdm

from datasets import load_dataset

import torch

from template import (preprocess_slimorca_mistral, preprocess_alpaca_mistral, preprocess_orcamath_mistral,
                      preprocess_slimorca_gemma, preprocess_alpaca_gemma, preprocess_orcamath_gemma
)

from transformers import AutoModelForCausalLM, AutoTokenizer

import argparse

tqdm.pandas()

def set_seed(seed_value=42):
    random.seed(seed_value)
    np.random.seed(seed_value)
    torch.manual_seed(seed_value)
    torch.cuda.manual_seed_all(seed_value)

def main(args):
    set_seed(42)

    model_name = args.model_path.split('/')[0]

    if model_name == 'mistralai':
        if args.data_name == "slimorca":
            dataset = load_dataset("Open-Orca/SlimOrca-Dedup", split='train')
            dataset = dataset.map(preprocess_slimorca_mistral)


        if args.data_name == "alpaca":
            dataset = load_dataset("tatsu-lab/alpaca", split='train')
            dataset = dataset.map(preprocess_alpaca_mistral)

        ## math
        if args.data_name == "orcamath":
            dataset = load_dataset("microsoft/orca-math-word-problems-200k", split='train')
            dataset = dataset.map(preprocess_orcamath_mistral)


    if model_name =='google':
        if args.data_name == "slimorca":
            dataset = load_dataset("Open-Orca/SlimOrca-Dedup", split='train')
            dataset = dataset.map(preprocess_slimorca_gemma)


        if args.data_name == "orcamath":
            dataset = load_dataset("microsoft/orca-math-word-problems-200k", split='train')
            dataset = dataset.map(preprocess_orcamath_gemma)

        if args.data_name == "alpaca":
            dataset = load_dataset("tatsu-lab/alpaca", split='train')
            dataset = dataset.map(preprocess_alpaca_gemma)


    tokenizer = AutoTokenizer.from_pretrained(args.model_path, trust_remote_code=True)
    answer_start = "### Assistant:"
    special_tokens = {"additional_special_tokens": [answer_start]}
    tokenizer.add_special_tokens(special_tokens)

    model = AutoModelForCausalLM.from_pretrained(
        args.model_path,
        output_attentions=True,
        torch_dtype=torch.bfloat16,
        device_map="auto",
        trust_remote_code=True,
    )

    model.resize_token_embeddings(len(tokenizer))

    answer_ids = tokenizer(answer_start)['input_ids'][1]

    def calculate_token_length(text):
        return {'token_length': len(tokenizer.tokenize(text['text']))}

    dataset = dataset.map(calculate_token_length)
    dataset = dataset.filter(lambda example: example['token_length'] <= args.max_length)


    bar = tqdm(enumerate(dataset), total=len(dataset))

    def calc_data(text):
        inputs = tokenizer(text, return_tensors='pt')

        answer_start_idx = (inputs['input_ids'] == answer_ids).nonzero(as_tuple=True)[1]

        if answer_start_idx.nelement() != 0:
            answer_start_idx = answer_start_idx[0] + 1
            extracted_values = inputs['input_ids'][0, answer_start_idx:]
        else:
            extracted_values = torch.tensor([], device=model.device)

        labels_length = inputs['input_ids'].size(1)
        labels = torch.full((1, labels_length), -100, device=model.device)

        labels[0, :extracted_values.size(0)] = extracted_values

        inputs['labels'] = labels

        inputs = {k: v.to(model.device) for k, v in inputs.items()}

        with torch.no_grad():
            logits = model(**inputs)


        attentions = logits.attentions

        std_devs_per_layer = []
        for layer in range(len(attentions)):
            layer_attentions = attentions[layer]
            std_devs_per_token = []

            for token_idx in range(answer_start_idx, inputs['input_ids'].size(1)):
                attention_scores = layer_attentions[0, :, token_idx, :token_idx]
                std_dev = torch.std(attention_scores, dim=-1).mean().item()
                std_devs_per_token.append(std_dev)

            std_devs_per_layer.append(torch.mean(torch.tensor(std_devs_per_token)).item())

        mean_std_dev_across_layers = torch.mean(torch.tensor(std_devs_per_layer)).item()

        return logits.loss.item(), mean_std_dev_across_layers


    loss_list = []
    attention_list = []

    for i, data in bar:
        loss, attention = calc_data(data['text'])
        loss_list.append(loss)
        attention_list.append(attention)

    df = dataset.to_pandas()
    df['loss'] = loss_list
    df['attention'] = attention_list

    data_name = args.model_path.split('/')[0] +  '_' + args.data_name

    df.to_parquet(f'{data_name}.parquet', index=False)


if __name__ == '__main__':
    parser = argparse.ArgumentParser()

    parser.add_argument('--model_path', type=str, default='mistralai/Mistral-7B-v0.1')
    parser.add_argument('--data_name', type=str, default='alpaca')
    parser.add_argument('--max_length', type=int, default=1024)

    args = parser.parse_args()
    main(args)