Mvanypersele/luciole_RAG

RAG citation benchmarks

Prompt-agnostic RAG benchmark exports for citation and retrieval-grounded QA evaluation.

Each row has:

• id: stable example id
• query: user question
• retrieved_documents: retrieved context chunks
• titles: chunk titles aligned with retrieved_documents
• supporting_index: zero-based indices of gold/supporting chunks
• answer: expected answer; an empty answer with empty supports represents an unanswerable row

The files are grouped as Hugging Face dataset configurations by source dataset.

Configs and splits

hotpotqa

split	file	rows	size
train	hotpotqa/train.jsonl	90447	540309604 bytes
validation	hotpotqa/validation.jsonl	7405	44653462 bytes

hotpotqa_fr

split	file	rows	size
train	hotpotqa_fr/train.jsonl	5007	52482039 bytes
test	hotpotqa_fr/test.jsonl	500	5178833 bytes

newsquadfr

split	file	rows	size
train	newsquadfr/train.jsonl	1650	14380040 bytes
validation	newsquadfr/validation.jsonl	455	4023374 bytes
test	newsquadfr/test.jsonl	415	3552859 bytes

piaf

split	file	rows	size
train	piaf/train.jsonl	8733	71794247 bytes
test	piaf/test.jsonl	491	4031761 bytes

squad2_fr_pragnakalp

split	file	rows	size
train	squad2_fr_pragnakalp/train.jsonl	78144	764627411 bytes
test	squad2_fr_pragnakalp/test.jsonl	925	9332188 bytes

tatqa

split	file	rows	size
train	tatqa/train.jsonl	13215	88719321 bytes
validation	tatqa/dev.jsonl	1668	11763957 bytes
test	tatqa/test.jsonl	1663	11189747 bytes

hotpot_native_french

split	file	rows	size
test	hotpot_native_french/test.jsonl	529	4714819 bytes

French multi-hop questions written directly on French Wikipedia, not translated. Each question is a bridge question over two abstracts: the first names an entity that leads to the second, which holds the answer. retrieved_documents holds ten abstracts, the two supporting ones and eight distractors, in their original order: the supporting abstracts are not placed first, supporting_index gives their positions. titles are the Wikipedia article titles.

The rows come from the test split of the filtered_qwen_v2 configuration of Mvanypersele/hotpotqa-fr-abstracts (956 questions, snapshot 20240916.fr). Only the questions judged valid in a review of the whole split on six criteria were kept: correct answer, bridge established by the first abstract, answer supported by the second, both abstracts needed, clear question, answer of the right type and scope. 568 of the 956 passed. A second check then removed 39 questions that designate an entity by a vague word or relation instead of what it is (« la propriété » for a chimney, « la ville liée à la mort de… »), leaving 529. The texts derive from French Wikipedia and are distributed under CC BY-SA 4.0.

openrag_bench

split	file	rows	size
test	openrag_bench/test.jsonl	686	14356762 bytes

Converted from the reranked_gold.parquet export of Mvanypersele/openrag_bench. Each row holds ten reranked chunks from long documents (financebench 10-K/10-Q filings, CUAD contracts, presentations, meeting transcripts...). titles are the source filenames, with a [n] suffix when a file appears several times in the same row. supporting_index covers the chunks marked gold, gold_alternative or gold_equivalent, so a row can have several supports. 97 rows are naturally unanswerable (answerable == False in the source): their answer and supporting_index are empty.