mmiakashs's picture
Release dataset generator
165da3c verified
# Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
# SPDX-License-Identifier: CC-BY-NC-4.0
from typing import List, Dict
import uuid
from loguru import logger
from models import DocumentAsset, SplicedDocument, SourceDocument, GroundTruthPage
from services.shuffle_strategies.base_strategy import BaseStrategy
class MonoSeq(BaseStrategy):
"""DocSplit-Mono-Seq: Single category document concatenation sequentially.
Creates document packets by concatenating entire documents from the same category
while preserving original page order. Tests boundary detection without
category transitions as discriminative signals.
"""
def generate(
self,
documents_by_type: Dict[str, List[DocumentAsset]],
doc_names_for_split: Dict[str, List[str]],
num_spliced_docs: int
) -> List[SplicedDocument]:
available_docs = self._get_available_docs(documents_by_type, doc_names_for_split)
# Filter out language category for large datasets (min_pages >= 20)
if self.min_pages >= 20 and "language" in available_docs:
del available_docs["language"]
spliced_documents = []
for i in range(num_spliced_docs):
# Pick a random document type
doc_type = self.rng.choice(list(available_docs.keys()))
# Set target page count
target_pages = self.rng.randint(self.min_pages, self.max_pages)
# Keep adding entire documents until reaching target
source_documents = []
ground_truth = []
current_page = 1
used_docs = set()
while current_page - 1 < target_pages:
# Get available docs of this type that haven't been used
available = [d for d in available_docs[doc_type]
if d.doc_name not in used_docs and d.page_count <= self.max_pages]
if not available:
break
# Pick a random document
doc = self.rng.choice(available)
used_docs.add(doc.doc_name)
# Check if adding this doc would exceed max_pages
if current_page - 1 + doc.page_count > self.max_pages:
continue
# Add all pages from this document in original order
pages = list(range(1, doc.page_count + 1))
source_documents.append(SourceDocument(
doc_type=doc.doc_type,
doc_name=doc.doc_name,
pages=pages
))
# Add ground truth for each page
for source_page in pages:
ground_truth.append(GroundTruthPage(
page_num=current_page,
doc_type=doc.doc_type,
source_doc=doc.doc_name,
source_page=source_page
))
current_page += 1
# Only add if we have at least min_pages
if current_page - 1 >= self.min_pages:
spliced_doc = SplicedDocument(
spliced_doc_id=str(uuid.uuid4()),
source_documents=source_documents,
ground_truth=ground_truth,
total_pages=current_page - 1
)
spliced_documents.append(spliced_doc)
logger.info(f"Generated {len(spliced_documents)} DocSplit-Mono-Seq documents")
return spliced_documents