mattmorgis/bioasq-12b-rag

BioASQ 12B RAG Dataset

A processed version of the BioASQ 12B dataset optimized for Retrieval-Augmented Generation (RAG) applications in biomedical question answering.

This dataset contains two distinct subsets specifically designed for RAG applications:

1. A text corpus of PubMed abstracts ready for indexing and retrieval, containing detailed metadata and full abstract text.

2. An evaluation dataset consisting of biomedical questions, each paired with an ideal answer and a list of passage IDs that are relevant to answering the question.

This structure makes it ideal for building and evaluating RAG systems that retrieve relevant biomedical information from a corpus and generate accurate, evidence-based answers to complex biomedical questions.

The code to generate this dataset is here: https://github.com/MattMorgis/bioasq-rag

Dataset Structure

The dataset contains three main components:

1. Corpus (data/corpus.jsonl): A collection of PubMed abstracts including metadata.

   • The corpus is accessible through the "train" split of the "text-corpus" config
   • Each document contains:
     • id: PubMed ID
     • title: Title of the paper
     • text: Abstract text
     • url: PubMed URL
     • publication_date: Publication date
     • journal: Journal name
     • authors: List of authors
     • doi: Digital Object Identifier (if available)
     • keywords: Keywords
     • mesh_terms: MeSH terms

2. Dev Questions (data/dev.jsonl): Development set of biomedical questions.

   • The dev questions are accessible through the "dev" split of the "question-answer-passages" config
   • Each question contains:
     • question_id: Unique identifier for the question
     • question: The question text
     • answer: Ideal answer
     • relevant_passage_ids: List of PubMed IDs for relevant abstracts
     • type: Question type (e.g., factoid, list, yes/no, summary)
     • snippets: Relevant snippets from abstracts

3. Eval Questions (data/eval.jsonl): Eval set of biomedical questions.

   • Same structure as dev questions, accessible through the "eval" split

Usage

This dataset is designed for training and evaluating RAG systems for biomedical question answering.

Loading the Dataset

You can load the dataset using the Hugging Face datasets library. Note that you must specify a config name:

from datasets import load_dataset

# Load the corpus of PubMed abstracts
corpus_dataset = load_dataset("mattmorgis/bioasq-12b-rag", "text-corpus")

# Load the question-answer dataset
questions_dataset = load_dataset("mattmorgis/bioasq-12b-rag", "question-answer-passages")

# Access the corpus data (note: the corpus is stored in the "train" split)
corpus_docs = corpus_dataset["train"]

# Access the development questions
dev_questions = questions_dataset["dev"]

# Access the eval questions
eval_questions = questions_dataset["eval"]

Example RAG Application

This dataset can be used to build a biomedical RAG system:

1. Index the corpus using a vector database (e.g., FAISS, Chroma)
2. Embed questions using a biomedical or general purpose text embedding model
3. Retrieve relevant documents from the corpus based on question embeddings
4. Generate answers using a large language model (LLM) with the retrieved context

Evaluation

The dataset provides gold standard answers and relevant passage IDs that can be used to evaluate:

• Retrieval accuracy
• Answer quality
• Domain-specific knowledge incorporation

Source

This dataset is derived from the BioASQ Challenge task 12b dataset.

Anastasia Krithara, Anastasios Nentidis, Konstantinos Bougiatiotis, Georgios Paliouras. BioASQ-QA: A manually curated corpus for Biomedical Question Answering. bioRxiv 2022.12.14.520213; doi: https://doi.org/10.1101/2022.12.14.520213