File size: 7,304 Bytes
4c30587
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
"""
LongMemEval dataset processor.

LongMemEval is a long-context memory evaluation dataset with chat history sessions.
Each sample contains multiple haystack sessions (chat history) and a question-answer pair.
"""
from typing import List, Dict, Any
from tqdm import tqdm
import tiktoken

from .base import DataProcessor, ChunkMode, register_processor


@register_processor("longmemeval")
class LongMemEvalProcessor(DataProcessor):
    """
    Data processor for LongMemEval dataset.

    LongMemEval format:
    {
        "question": "...",
        "answer": "...",
        "question_date": "...",
        "haystack_sessions": [
            [{"role": "user/assistant", "content": "..."}...],
            ...
        ],
        "haystack_dates": ["date1", "date2", ...]
    }
    """

    supported_chunk_modes = [
        ChunkMode.TURN,
        ChunkMode.TURN_PAIR,
        ChunkMode.FULL_SESSION,
        ChunkMode.FIXED_LENGTH
    ]

    def __init__(
        self,
        chunk_mode: str = "turn-pair",
        chunk_size: int = 1024,
        chunk_overlap: int = 128,
        tokenizer_name: str = "cl100k_base",
        show_progress: bool = False
    ):
        """
        Initialize LongMemEval processor.

        Args:
            chunk_mode: How to split chat history into chunks
            chunk_size: Target size of each chunk in tokens (fixed-length mode)
            chunk_overlap: Number of overlapping tokens between chunks
            tokenizer_name: Tiktoken tokenizer name for token counting
            show_progress: Whether to show progress bars during processing
        """
        super().__init__(chunk_mode)
        self.chunk_size = chunk_size
        self.chunk_overlap = chunk_overlap
        self.show_progress = show_progress
        try:
            self.tokenizer = tiktoken.get_encoding(tokenizer_name)
        except Exception:
            self.tokenizer = tiktoken.get_encoding("cl100k_base")

    def extract_chunks(self, data: Dict) -> List[str]:
        """
        Extract text chunks from LongMemEval haystack sessions.

        Args:
            data: LongMemEval data dict containing 'haystack_sessions'

        Returns:
            List of text chunks with optional date prefixes
        """
        chunks = []
        haystack_sessions = data.get('haystack_sessions', [])
        haystack_dates = data.get('haystack_dates', [''] * len(haystack_sessions))

        session_iter = zip(haystack_sessions, haystack_dates)
        if self.show_progress:
            session_iter = tqdm(list(session_iter), desc='Extracting sessions', leave=False)

        for session_entry, session_date in session_iter:
            if not isinstance(session_entry, list):
                continue

            session_chunks = self._extract_session_chunks(
                session_entry, session_date
            )
            chunks.extend(session_chunks)

        return chunks

    def _extract_session_chunks(self, turns: List[Dict],
                                date_str: str) -> List[str]:
        """
        Extract chunks from a single session based on chunk mode.

        Args:
            turns: List of turn dicts with 'role' and 'content'
            date_str: Date string for this session

        Returns:
            List of text chunks
        """
        chunks = []

        if self.chunk_mode == ChunkMode.TURN:
            for turn in turns:
                text = f"{turn['role']}: {turn['content']}"
                if date_str:
                    text = f"Session Date: {date_str}\n{text}"
                chunks.append(text)

        elif self.chunk_mode == ChunkMode.TURN_PAIR:
            for j in range(0, len(turns), 2):
                pair_texts = []
                for k in range(j, min(j + 2, len(turns))):
                    pair_texts.append(f"{turns[k]['role']}: {turns[k]['content']}")

                text = "\n\n".join(pair_texts)
                if date_str:
                    text = f"Session Date: {date_str}\n\n{text}"
                chunks.append(text)

        elif self.chunk_mode == ChunkMode.FULL_SESSION:
            session_text = "\n\n".join([
                f"{turn['role']}: {turn['content']}"
                for turn in turns
            ])
            if date_str:
                session_text = f"Session Date: {date_str}\n\n{session_text}"
            chunks.append(session_text)
        elif self.chunk_mode == ChunkMode.FIXED_LENGTH:
            session_text = "\n\n".join([
                f"{turn['role']}: {turn['content']}"
                for turn in turns
            ])
            fixed_chunks = self._split_fixed_length(session_text)
            for chunk in fixed_chunks:
                text = chunk
                if date_str:
                    text = f"Session Date: {date_str}\n\n{text}"
                chunks.append(text)

        return chunks

    def _split_fixed_length(self, text: str) -> List[str]:
        """
        Split text into fixed-length chunks based on token count.

        Args:
            text: Input text to split

        Returns:
            List of text chunks, each approximately chunk_size tokens
        """
        if not text:
            return []

        tokens = self.tokenizer.encode(text)
        if len(tokens) <= self.chunk_size:
            return [text]

        overlap = min(self.chunk_overlap, max(self.chunk_size - 1, 0))
        chunks = []
        start = 0
        while start < len(tokens):
            end = start + self.chunk_size
            chunk_tokens = tokens[start:end]
            chunk_text = self.tokenizer.decode(chunk_tokens)
            chunk_text = chunk_text.strip()
            if chunk_text:
                chunks.append(chunk_text)
            start = end - overlap
            if start >= len(tokens) - overlap:
                break
        return chunks

    def get_sample_id(self, data: Dict) -> str:
        """Get sample ID from LongMemEval data."""
        question_id = data.get('question_id')
        if question_id:
            return str(question_id)
        # LongMemEval may not have explicit sample_id, generate from question hash
        if 'sample_id' in data:
            return data['sample_id']
        question = data.get('question', '')
        return f"lme_{hash(question) % 10**8}"

    def get_qa_list(self, data: Dict) -> List[Dict[str, Any]]:
        """
        Extract QA list from LongMemEval data.

        LongMemEval has a single question-answer pair per sample.

        Returns:
            List with single QA dict containing:
            - question: str
            - answer: str
            - question_date: str (optional)
        """
        if 'question' not in data or 'answer' not in data:
            return []

        return [{
            'question': data['question'],
            'answer': data['answer'],
            'question_date': data.get('question_date', '')
        }]

    def get_metadata(self, data: Dict) -> Dict[str, Any]:
        """
        Extract additional metadata.

        Returns:
            Dict with session count and other info
        """
        haystack_sessions = data.get('haystack_sessions', [])
        return {
            'num_sessions': len(haystack_sessions),
            'question_date': data.get('question_date', '')
        }