LOFT RAG Datasets

This is the main index for all LOFT (Long-context Open Foundation Tasks) RAG (Retrieval-Augmented Generation) datasets.

Overview

All datasets are part of the LOFT benchmark and have been converted to HuggingFace format with 100% prompt fidelity to the original LOFT implementation.

Available Datasets

HotpotQA

• 128k: loft-rag-hotpotqa-128k
• 1m: loft-rag-hotpotqa-1m
• 32k: loft-rag-hotpotqa-32k

MuSiQue

• 128k: loft-rag-musique-128k
• 1m: loft-rag-musique-1m
• 32k: loft-rag-musique-32k

Natural Questions

• 128k: loft-rag-nq-128k
• 1m: loft-rag-nq-1m
• 32k: loft-rag-nq-32k

Qampari

• 128k: loft-rag-qampari-128k
• 1m: loft-rag-qampari-1m
• 32k: loft-rag-qampari-32k

Quest

• 128k: loft-rag-quest-128k
• 1m: loft-rag-quest-1m
• 32k: loft-rag-quest-32k

Usage

from datasets import load_dataset

# Load any dataset
dataset = load_dataset("loft-rag-nq-32k")

# Access splits
dev_data = dataset["dev"]
test_data = dataset["test"]

# Convert to pandas
df_dev = dev_data.to_pandas()
df_test = test_data.to_pandas()

Dataset Structure

All datasets contain:

• context: Full prompt context with corpus documents and few-shot examples
• question: Query separator + query format + query text
• answer_prefix: Prefix for answer generation
• answers: Ground truth answers (list)
• task: Task identifier
• max_new_tokens: Maximum tokens for generation (256)

Citation

@article{loft2024,
  title={LOFT: Long-context Open Foundation Tasks},
  author={Google DeepMind},
  year={2024},
  url={https://github.com/google-deepmind/loft}
}

License

Apache 2.0