tiktoc/data_preprocessing.py at main · umass-ml4ed/tiktoc · 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
161
162
163
164
165
166
167
import pandas as pd
import numpy as np
import torch
import hydra
from model import *
from pdb import set_trace
import pickle


def nested_dict():
    return defaultdict(list)

# dataset.pkl contains GPT-2 question representation embedding only
# Saved dataset_wo_emb.pkl to create embedding based on LLM used in future work.
def clean_dataset(configs):
    dataset = pd.read_pickle(configs.data_path + '/dataset.pkl')
    dataset = dataset.drop(columns=['prompt-embedding', 'input'])
    dataset.to_pickle(configs.data_path + '/dataset_wo_emb.pkl')


# There are 50 unique programming questions.
# Returns a dictionary (key: question prompt, value: question representation embedding)
def get_questions_prompt(configs, tokenizer, model, device):
    dataset = pd.read_pickle(configs.data_path + '/dataset.pkl')
    questions = list(dataset['prompt'].unique())

    question_embed_dict = {}

    for q_i in questions:
        prompt_embedding = tokenizer(q_i, return_tensors='pt')

        if configs.okt_model != 'codellama/CodeLlama-7b-Instruct-hf' and configs.okt_model != 'meta-llama/Meta-Llama-3-8B-Instruct' and configs.okt_model != 'Qwen/Qwen1.5-7B':
            token_embedding = model.transformer.wte(prompt_embedding['input_ids'].to(device))
        else:
            token_embedding = model.base_model.model.model.embed_tokens(prompt_embedding['input_ids'].to(device))

        avg_token_embedding = torch.mean(token_embedding[0], axis=0)
        question_embed_dict[q_i] = avg_token_embedding.cpu().detach().numpy()

    return question_embed_dict

def get_full_test_question_emb(question_embed_dict):
    prompt_df = pd.read_csv(os.path.join('test-case-query-results/prompt_concept_summary.csv'), on_bad_lines='warn')
    all_good = prompt_df.loc[prompt_df['Test Case Status'] == 'All Good']
    sat_questions = all_good['Requirement'].tolist()

    sat_questions_dict = {}
    for question in sat_questions:
        sat_questions_dict[question] = question_embed_dict[question].tolist()

    return sat_questions_dict


# Construct the input used in knowledge representation model
# Returns the dataset with input embedding, dimension should be 968(768+200) or 4296(4096+200)
def construct_dataset(configs, tokenizer, model, device):
    question_embed_dict = get_questions_prompt(configs, tokenizer, model, device)
    dataset = pd.read_pickle(configs.data_path + '/dataset_wo_emb_time.pkl')

    dataset['prompt-embedding'] = dataset['prompt'].map(question_embed_dict)
    dataset['input'] = dataset[['prompt-embedding', 'embedding']].apply(lambda x: np.concatenate((x[0], x[1])), axis=1)
    dataset['input'] = dataset['input'].apply(lambda x: torch.from_numpy(x))

    sat_questions = get_full_test_question_emb(question_embed_dict)

    return dataset, sat_questions

def handle_score_mismatch(configs):
    gt_code_ls = []
    student_ls = []
    prompt_ls = []
    score_ls = []
    bi_score_ls = []

    for i in range(7):
        cnt = 0
        filename = '/student_pair_all_' + str(i) + '.pkl'
        dataset = pd.read_pickle(configs.data_path + filename)

        for student, info in dataset.items():
            codes = info['ground_truth_codes']
            bi_scores = info['binary_correctness']
            match_ls = info['score_match']
            prompts = info['prompts']
            scores = info['scores']

            for code, bi_score, match, prompt, score in zip(codes, bi_scores, match_ls, prompts, scores):
                    if not match:
                        cnt += 1
                        gt_code_ls.append(code)
                        student_ls.append(student)
                        prompt_ls.append(prompt)
                        score_ls.append(score)
                        bi_score_ls.append(bi_score)

        print('mismatch count:', cnt)

    final_dict = {'ground_truth_codes': gt_code_ls, 'prompts': prompt_ls, 'students': student_ls, 'scores': score_ls, 'binary_correctness': bi_score_ls}

    with open(os.path.join(configs.data_path, 'mismatch_subset.pkl'), 'wb') as f:
            pickle.dump(final_dict, f)

    df = pd.DataFrame(final_dict)

    pa = configs.data_path+ '/mismatch_df.csv'
    df.to_csv(pa)


def preprocess_all_submission(configs, orig_df):
    gt_code_ls = []
    student_ls = []
    prompt_ls = []
    score_ls = []
    bi_score_ls = []

    for i in range(7):
        filename = '/student_pair_all_' + str(i) + '.pkl'
        dataset = pd.read_pickle(configs.data_path + filename)

        for student, info in dataset.items():
            codes = info['ground_truth_codes']
            bi_scores = info['binary_correctness']
            match_ls = info['score_match']
            prompts = info['prompts']
            scores = info['scores']

            for code, bi_score, match, prompt, score in zip(codes, bi_scores, match_ls, prompts, scores):
                    if match:
                        gt_code_ls.append(code)
                        student_ls.append(student)
                        prompt_ls.append(prompt)
                        score_ls.append(score)
                        bi_score_ls.append(bi_score)


    final_dict = {'Code': gt_code_ls, 'prompt': prompt_ls, 'SubjectID': student_ls, 'Score': score_ls, 'binary_correctness': bi_score_ls}

    df = pd.DataFrame(final_dict)
    # df.to_csv(configs.data_path + '/all_submission_preprocessed.csv')

    combined_df = pd.merge(orig_df, df, on=['SubjectID', 'Code', 'Score'], how='right')
    combined_df = combined_df.drop_duplicates(subset=['SubjectID', 'Code', 'timestep'], keep='first')

    combined_df.to_pickle(configs.data_path + '/dataset_granular_all.pkl')


def match_timestep(configs, tokenizer, model, device):
    dataset, sat_questions = construct_dataset(configs, tokenizer, model, device)

    maintable = pd.read_csv(configs.data_path + '/MainTable.csv')

    filtered = maintable[maintable['EventType'] == 'Run.Program']
    filtered = filtered.drop(columns=['Order', 'ToolInstances','ServerTimezone', 'CourseID','CourseSectionID', 'Score', 'Compile.Result', 'CompileMessageType', 'CompileMessageData', 'EventID', 'ParentEventID', 'SourceLocation', 'IsEventOrderingConsistent', 'EventType'])

    combined_df = pd.merge(dataset, filtered, on=['SubjectID', 'AssignmentID', 'ProblemID', 'CodeStateID'], how='left')

    combined_df = combined_df.sort_values(by=["SubjectID", "ServerTimestamp"])
    combined_df.to_pickle(configs.data_path + '/dataset_time.pickle')


@hydra.main(version_base=None, config_path=".", config_name="configs_okt")
def main(configs):
    preprocess_all_submission(configs, [])

if __name__ == "__main__":
    main()