Overview

Sencha is designed to test how well chunking algorithms handle long-form scientific documents. It contains full-text NLP research papers with questions that require finding specific information across multiple sections.

Key Challenges

• Handling structured sections (Abstract, Methods, Results, etc.)
• Preserving citation context (BIBREF tags)
• Managing hierarchical section headers
• Chunking technical content with equations and terminology

Dataset Structure

Corpus

The corpus config contains 250 full-text NLP papers.

Questions

The questions config contains 1,146 questions about paper content.

Statistics

Usage

from datasets import load_dataset

# Load the corpus
corpus = load_dataset("chonkie-ai/sencha", "corpus", split="train")

# Load the questions
questions = load_dataset("chonkie-ai/sencha", "questions", split="train")

# Use with MTCB evaluator
from mtcb import SenchaEvaluator
from chonkie import RecursiveChunker

evaluator = SenchaEvaluator(
    chunker=RecursiveChunker(chunk_size=512),
    embedding_model="voyage-3-large"
)
result = evaluator.evaluate(k=[1, 3, 5, 10])

Sample Topics

The papers cover various NLP topics including:

• Sentiment analysis and affective computing
• Word embeddings and language models
• Text classification and NER
• Question answering systems
• Machine translation
• Social media analysis
• Clinical NLP

Source

Derived from QASPER (NAACL 2021) by Allen AI - a dataset for question answering on scientific research papers.

License

CC-BY-4.0 (following QASPER license)