File size: 6,089 Bytes
f50dc54
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
TTRLVR Dataset for AZR Integration

TTRLVR parquet ํŒŒ์ผ์„ ์ฝ์–ด AZR ํ•™์Šต์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ๋ณ€ํ™˜
"""

import os
import pandas as pd
import numpy as np
from typing import Dict, List, Any, Optional
from torch.utils.data import Dataset
from transformers import AutoTokenizer

from .rl_dataset import RLHFDataset


class TTRLVRDataset(RLHFDataset):
    """TTRLVR ๋ฐ์ดํ„ฐ๋ฅผ AZR ํ˜•์‹์œผ๋กœ ๋กœ๋“œํ•˜๋Š” Dataset"""
    
    def __init__(self, 
                 parquet_files: str,
                 tokenizer: AutoTokenizer,
                 task_type: Optional[str] = None,
                 **kwargs):
        """
        Args:
            parquet_files: TTRLVR parquet ํŒŒ์ผ ๊ฒฝ๋กœ (๋˜๋Š” ๋””๋ ‰ํ† ๋ฆฌ)
            tokenizer: ํ† ํฌ๋‚˜์ด์ €
            task_type: ํŠน์ • task ํƒ€์ž…๋งŒ ๋กœ๋“œ (induction/deduction/abduction)
        """
        # parquet_files๊ฐ€ ListConfig์ธ ๊ฒฝ์šฐ ๋ฆฌ์ŠคํŠธ๋กœ ๋ณ€ํ™˜
        from omegaconf import ListConfig
        if isinstance(parquet_files, ListConfig):
            parquet_files = list(parquet_files)
        
        # parquet_files๊ฐ€ ๋””๋ ‰ํ† ๋ฆฌ์ธ ๊ฒฝ์šฐ ์ฒ˜๋ฆฌ
        if isinstance(parquet_files, str) and os.path.isdir(parquet_files):
            if task_type:
                parquet_files = os.path.join(parquet_files, f"{task_type}.parquet")
            else:
                # ๋ชจ๋“  task ํƒ€์ž… ํŒŒ์ผ ์ˆ˜์ง‘
                files = []
                for t in ['induction', 'deduction', 'abduction']:
                    f = os.path.join(parquet_files, f"{t}.parquet")
                    if os.path.exists(f):
                        files.append(f)
                parquet_files = files
                
        super().__init__(
            parquet_files=parquet_files,
            tokenizer=tokenizer,
            prompt_key='prompt',  # TTRLVR์€ 'prompt' ํ‚ค ์‚ฌ์šฉ
            **kwargs
        )
        
        self.task_type = task_type
        
    def __getitem__(self, idx):
        """๋‹จ์ผ ์ƒ˜ํ”Œ ๋ฐ˜ํ™˜"""
        # TTRLVR ํŠน๋ณ„ ์ฒ˜๋ฆฌ - ์›๋ณธ ๋ฐ์ดํ„ฐ์— ๋จผ์ € ์ ‘๊ทผ
        # RLHFDataset.__getitem__์ด prompt๋ฅผ popํ•˜๊ธฐ ์ „์— ์›๋ณธ ๋ฐ์ดํ„ฐ ๊ฐ€์ ธ์˜ค๊ธฐ
        if hasattr(self, 'dataframe') and hasattr(self.dataframe, 'iloc'):
            # pandas DataFrame์ธ ๊ฒฝ์šฐ
            original_row = self.dataframe.iloc[idx].to_dict()
            # prompt ํ•„๋“œ ๋ฐฑ์—…
            original_prompt = original_row.get('prompt', None)
        else:
            original_row = {}
            original_prompt = None
        
        # ๊ธฐ๋ณธ ๋ฐ์ดํ„ฐ ๊ฐ€์ ธ์˜ค๊ธฐ (RLHFDataset.__getitem__ ํ˜ธ์ถœ)
        data = super().__getitem__(idx)
        
        # ๋ฐฑ์—…ํ•œ prompt ์‚ฌ์šฉ
        row = original_row
        
        # prompt ์ฒ˜๋ฆฌ - numpy array, list, dict ๋“ฑ ๋‹ค์–‘ํ•œ ํ˜•ํƒœ ์ฒ˜๋ฆฌ
        prompt_text = None
        if original_prompt is not None:
            if isinstance(original_prompt, np.ndarray):
                if len(original_prompt) > 0 and isinstance(original_prompt[0], dict):
                    prompt_text = original_prompt[0].get('content', '')
                else:
                    prompt_text = str(original_prompt)
            elif isinstance(original_prompt, list):
                if len(original_prompt) > 0 and isinstance(original_prompt[0], dict):
                    prompt_text = original_prompt[0].get('content', '')
                else:
                    prompt_text = str(original_prompt)
            elif isinstance(original_prompt, str):
                prompt_text = original_prompt
            else:
                prompt_text = str(original_prompt)
        
        # prompt๋ฅผ data์— ์ถ”๊ฐ€ (๋ฌธ์ž์—ด๋กœ)
        if prompt_text is not None:
            data['prompt'] = prompt_text
        
        # TTRLVR ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ ์ถ”๊ฐ€
        ttrlvr_metadata = {
            'task_type': self._extract_task_type(row),
            'expected_solution': row.get('ground_truth', ''),
            'problem': row.get('problem', {}),
            'ipo_group_id': row.get('ipo_group_id', ''),
            'uid': row.get('uid', ''),
            'evaluation_data': self._prepare_evaluation_data(row)
        }
        
        # ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋ฅผ data์— ์ถ”๊ฐ€
        data['ttrlvr_metadata'] = ttrlvr_metadata
        
        return data
    
    def _extract_task_type(self, row: pd.Series) -> str:
        """ํ–‰์—์„œ task ํƒ€์ž… ์ถ”์ถœ"""
        uid = row.get('uid', '')
        if 'induction' in uid:
            return 'induction'
        elif 'deduction' in uid:
            return 'deduction'
        elif 'abduction' in uid:
            return 'abduction'
        return 'unknown'
    
    def _prepare_evaluation_data(self, row: pd.Series) -> Dict[str, Any]:
        """Task ํƒ€์ž…๋ณ„ evaluation data ์ค€๋น„"""
        
        # 1. ๋จผ์ € ์ €์žฅ๋œ evaluation_data๊ฐ€ ์žˆ๋Š”์ง€ ํ™•์ธ (Phase 1-4์—์„œ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ)
        if 'evaluation_data' in row and row['evaluation_data']:
            eval_data = row['evaluation_data']
            # pandas/numpy ๊ฐ์ฒด๋ฅผ Python ๋„ค์ดํ‹ฐ๋ธŒ ํƒ€์ž…์œผ๋กœ ๋ณ€ํ™˜
            if hasattr(eval_data, 'item'):
                eval_data = eval_data.item()
            return eval_data if isinstance(eval_data, dict) else {}
        
        # 2. Fallback: problem ํ•„๋“œ์—์„œ ๊ตฌ์„ฑ (ํ˜ธํ™˜์„ฑ ์œ ์ง€)
        task_type = self._extract_task_type(row)
        problem = row.get('problem', {})
        
        if task_type == 'induction':
            # IPO์—์„œ input/output ์Œ ์ถ”์ถœ
            return {
                'input_output_pairs': [
                    (problem.get('input', ''), 
                     problem.get('output', ''))
                ]
            }
        elif task_type == 'deduction':
            return {
                'function_code': problem.get('snippet', ''),
                'input': problem.get('input', '')
            }
        elif task_type == 'abduction':
            return {
                'function_code': problem.get('snippet', ''),
                'expected_output': problem.get('output', '')
            }
        
        return {}