f20180301/loft-rag-nq-1m

LOFT RAG - Natural Questions (1m)

Dataset Description

This dataset is part of the LOFT (Long-context Open Foundation Tasks) benchmark, specifically the RAG (Retrieval-Augmented Generation) task.

• Dataset: Natural Questions
• Context Length: 1m
• Task Type: RAG (Retrieval-Augmented Generation)
• Language: English
• Source: LOFT Benchmark (Google DeepMind)

Dataset Structure

Data Fields

• context (string): Full prompt context including corpus documents and few-shot examples
• question (string): Query separator + query format + query text
• answer_prefix (string): Prefix for answer generation ("Final Answer: ")
• answers (list[string]): Ground truth answers
• task (string): Task identifier (e.g., "nq_1m")
• max_new_tokens (int64): Maximum tokens for generation (256)

Data Splits

• dev: Development set (10 examples)
• test: Test set (100 examples)

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("loft-rag-nq-1m")

# Access splits
dev_data = dataset["dev"]
df_dev = dev_data.to_pandas()
test_data = dataset["test"]
df_test = test_data.to_pandas()

# Example usage
sample = dataset["dev"][0] if "dev" in dataset else dataset["test"][0]
context = sample["context"]
question = sample["question"]
answers = sample["answers"]

Dataset Creation

This dataset was converted from LOFT's original format to HuggingFace format using exact LOFT prompt construction to ensure 100% fidelity.

• Prompt Construction: Uses LOFT's PromptRegistry and concatenate_chunks() for exact prompt matching
• Few-shot Examples: Preserved exactly as in LOFT (5 examples)
• Corpus Documents: Full corpus included in context (corpus-in-context approach)
• Verification: All prompts verified to match LOFT originals exactly

Related Datasets

All LOFT RAG datasets are available under the loft-rag-* namespace:

• Main Index - Overview of all datasets

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