gst-code-collection / MemSkill /src /eval /longmemeval.py
SitongGong's picture
merge code dirs 1031 (part 3)
4c30587 verified
Raw History Blame Contribute Delete
2.56 kB
"""
LongMemEval dataset evaluator.
Handles evaluation for LongMemEval dataset with its specific prompt format.
"""
from typing import List, Dict, Any, Tuple
from .base import Evaluator, register_evaluator
@register_evaluator("longmemeval")
class LongMemEvalEvaluator(Evaluator):
"""
Evaluator for LongMemEval dataset.
LongMemEval has a simpler structure with single question-answer pairs
and includes a question_date field for temporal context.
"""
def prepare_eval_args(self) -> Any:
"""Override default eval args to allow longer generations for HotpotQA."""
eval_args = super().prepare_eval_args()
eval_args.max_new_tokens = 512
return eval_args
def filter_qa_list(self, qa_list: List[Dict]) -> List[Tuple[int, Dict]]:
"""
Filter QA list. LongMemEval typically has all valid QA items.
Args:
qa_list: List of QA dicts
Returns:
List of (index, qa_dict) tuples for valid QA items
"""
valid_qa = []
for i, qa in enumerate(qa_list):
# Check that question and answer exist
if 'question' in qa and 'answer' in qa:
valid_qa.append((i, qa))
return valid_qa
def build_prompt(self, question: str, retrieved_memories: List[str],
qa_item: Dict) -> str:
"""
Build evaluation prompt for LongMemEval.
Args:
question: The question to answer
retrieved_memories: List of retrieved memory texts
qa_item: Original QA dict (contains question_date)
Returns:
Formatted prompt string
"""
from prompts.prompt_pool import LONGMEMEVAL_ANSWER_PROMPT
if len(retrieved_memories) > 0:
history_string = ""
for j, mem in enumerate(retrieved_memories):
history_string += f'\n### Session {j + 1}:\nSession Content:\n{mem}\n'
else:
history_string = "No relevant chat history available."
question_date = qa_item.get('question_date', '')
return LONGMEMEVAL_ANSWER_PROMPT.format(
history_string,
question_date,
question
)
def _get_result_metadata(self, qa: Dict) -> Dict[str, Any]:
"""
Extract LongMemEval-specific metadata.
Args:
qa: QA dict
Returns:
Metadata including question_date
"""
return {
'question_date': qa.get('question_date', '')
}