minkyungpark/flashrag_wiki18_aligned

FlashRAG Multi-hop QA — wiki-18 aligned gold_doc_ids

Multi-hop QA splits from FlashRAG (HotpotQA, MuSiQue, 2WikiMultiHopQA), extended with a per-sample list of paragraph chunk ids in PeterJinGo/wiki-18-corpus that contain the actual supporting evidence. All original FlashRAG fields are preserved verbatim; only one new field is added: metadata.gold_doc_ids: list[str].

Why this dataset exists

FlashRAG's multi-hop QA samples ship gold supervision at the Wikipedia article title level:

• HotpotQA / 2WikiMultiHopQA: metadata.supporting_facts.title + sent_id
• MuSiQue: metadata.question_decomposition[i].support_paragraph.title + paragraph_text

The standard retrieval corpus and index used by Search-R1, FlashRAG, and downstream multi-hop RAG research is PeterJinGo/wiki-18-corpus — Wikipedia split into 100-word paragraph chunks, where one article maps to ~5–10 chunks all sharing the same title field.

Concrete consequence: with title-only matching, every chunk of a supporting article is labeled gold even though only one paragraph actually contains the evidence. For example, the e5 retriever hosted over wiki-18 returns four Scott Derrickson chunks (ids 5441536, 5441537, 5441538, 14728950) in the top-5 of Scott Derrickson director. A title-based gold check marks all four as positive, inflating chunk-level metrics (gold-recall, write-precision) and adding label noise to chunk-level supervision (e.g., memory-write policies, retrieval re-ranking heads).

This dataset resolves the mismatch by aligning each supporting fact to the specific wiki-18 chunk(s) that actually contain the evidence sentence/paragraph.

How gold_doc_ids is computed

The alignment is deterministic and LLM-free: substring + n-gram matching against an in-memory title→chunks index built once over the entire wiki-18 corpus (~21M chunks, ~3.2M unique titles).

HotpotQA / 2WikiMultiHopQA

For each (supporting_title, sent_id):

1. Recover the literal sentence from metadata.context.sentences[idx][sent_id] (HotpotQA) / metadata.context.content[idx][sent_id] (2Wiki).
2. Look up wiki-18 chunks sharing that title (with (disambiguation) suffix-stripping fallback for titles like Méditerranée (1963 film) → Méditerranée).
3. A chunk is marked gold if either:
   • The normalized sentence is a substring of the chunk text, or
   • Word trigram recall (sentence trigrams covered by chunk trigrams) ≥ 0.6.

The trigram fallback absorbs minor wiki dump version drift: word-order shifts (Dutch-Belgian ↔ Belgian Dutch), single-word edits (English ↔ British), punctuation/quote-escape differences (""x"" ↔ "x").

Sentences that straddle a 100-word chunk boundary may produce more than one matching chunk; all matches are kept.

MuSiQue

For each hop's support_paragraph:

1. Look up wiki-18 chunks sharing the paragraph's title (same suffix fallback).
2. A chunk is marked gold if any of:
   • It is a substring of paragraph_text (or vice versa).
   • 5-gram word Jaccard between paragraph and chunk ≥ 0.5.
   • 3-gram recall (paragraph → chunk) ≥ 0.5.

Sample drop policy

A sample is kept only if every supporting hop aligns to at least one chunk. Samples where any hop fails (the article is missing from wiki-18, or no chunk passes substring / n-gram match due to wiki dump version drift) are dropped. No fallback to title-only labeling, no LLM relabeling — keeping label precision is the priority.

HotpotQA dev source

FlashRAG's hotpotqa/dev is the fullwiki variant: distractor pool is the retriever's top-10, so supporting articles are absent from the context for 64% of dev samples and the per-sentence text needed for substring matching is unavailable. HotpotQA dev in this dataset is therefore aligned from hotpotqa/hotpot_qa::distractor/validation — the same 7,405 official HotpotQA dev questions, with the same supporting_facts schema, but the distractor pool always contains the supporting articles. Train, MuSiQue, and 2WikiMultiHopQA use FlashRAG sources directly.

Final alignment statistics

Dataset	Split	Kept	Total	Rate	Source
hotpotqa	train	55,863	90,447	61.76%	RUC-NLPIR/FlashRAG_datasets::hotpotqa/train
hotpotqa	dev	4,559	7,405	61.57%	hotpotqa/hotpot_qa::distractor/validation
musique	train	9,952	19,938	49.91%	RUC-NLPIR/FlashRAG_datasets::musique/train
musique	dev	1,131	2,417	46.79%	RUC-NLPIR/FlashRAG_datasets::musique/dev
2wikimultihopqa	train	5,433	15,000	36.22%	RUC-NLPIR/FlashRAG_datasets::2wikimultihopqa/train
2wikimultihopqa	dev	5,704	12,576	45.36%	RUC-NLPIR/FlashRAG_datasets::2wikimultihopqa/dev

Drops are dominated by wiki-18 corpus coverage, not algorithm failures. For 2WikiMultiHopQA, ~36% of supporting articles are absent from wiki-18-corpus (long-tail entities — non-English-Wikipedia personalities, obscure films, etc. — that the 2018 Karpukhin DPR pipeline did not include). Multi-hop questions multiply the per-hop miss rate, which is why 2Wiki is lowest.

Usage

from datasets import load_dataset

ds = load_dataset("minkyungpark/flashrag_wiki18_aligned", "hotpotqa", split="train")
sample = ds[0]

# new field
sample["metadata"]["gold_doc_ids"]
# -> ["11260951", "11260954"]

# original FlashRAG fields are preserved unchanged
sample["question"]
sample["golden_answers"]
sample["metadata"]["supporting_facts"]
sample["metadata"]["context"]

gold_doc_ids strings are exact id values from PeterJinGo/wiki-18-corpus, matching the document.id field returned by retriever responses (e.g., e5-base-v2 + HNSW serving the same corpus). Verified 1:1 on actual retriever responses.

Long-context HotpotQA configs

Three long-context variants of hotpotqa (long-context-hotpotqa-{50,100,500}) are derived from the hotpotqa config above by inserting each sample's gold chunks into a context filled with random distractor chunks sampled from PeterJinGo/wiki-18-corpus. Construction follows the MemAgent recipe (random distractors, not retriever hard negatives), enabling long-context evaluation where the relevant chunks are guaranteed present amid varying amounts of noise.

Construction

For each sample in hotpotqa:

1. Look up the actual chunk text for every id in gold_doc_ids.
2. Sample N - len(gold_doc_ids) random chunks from the wiki-18 corpus, excluding the gold ids.
3. Shuffle the combined list with a per-sample seed (random.Random(f"{seed}:{sample_id}"), seed=4) so ordering is deterministic per sample but gold positions vary across samples.
4. Format as Document {i}:\n{title}\n{text} joined by \n\n.

The random distractor pool spans the full ~21M chunk corpus, not retriever top-K, matching MemAgent's setup where distractor relevance is uncorrelated with the query.

Schema

{
  "id":             str,
  "question":       str,
  "golden_answers": list[str],
  "gold_doc_ids":   list[str],       # chunk ids of gold docs (same as hotpotqa::metadata.gold_doc_ids)
  "doc_ids":        list[str],       # chunk ids of every doc in context, in Document-order (doc_ids[i] = "Document i+1")
  "gold_positions": list[int],       # 1-indexed Document numbers in context that are gold (e.g., [2, 3, 8, 10])
  "num_docs":       int,             # 50 | 100 | 500
  "context":        str              # "Document 1:\n{title}\n{text}\n\nDocument 2:\n..."
}

Usage

from datasets import load_dataset

# pick the document count you want
ds = load_dataset("minkyungpark/flashrag_wiki18_aligned", "long-context-hotpotqa-100", split="train")
sample = ds[0]

sample["question"]        # same as hotpotqa config
sample["golden_answers"]  # same as hotpotqa config
sample["gold_doc_ids"]    # same chunk ids as hotpotqa::metadata.gold_doc_ids
sample["doc_ids"]         # chunk ids of every doc in context, ordered by Document N
sample["gold_positions"]  # e.g., [2, 3, 8, 10] → Documents 2/3/8/10 are gold
sample["num_docs"]        # 100
sample["context"]         # gold + 96 random distractors, shuffled, formatted

Splits

Config	Split	Examples
long-context-hotpotqa-50	train	55,863
long-context-hotpotqa-50	dev	4,559
long-context-hotpotqa-100	train	55,863
long-context-hotpotqa-100	dev	4,559
long-context-hotpotqa-500	train	55,863
long-context-hotpotqa-500	dev	4,559

Schema

Row keys are identical to FlashRAG; gold_doc_ids is added inside metadata:

{
  "id":             str,
  "question":       str,
  "golden_answers": list[str],
  "metadata": {
    "gold_doc_ids": list[str],   # NEW — wiki-18 chunk ids
    ... (all original fields preserved verbatim, see dataset_info above)
  }
}

Sources

• Retrieval corpus: PeterJinGo/wiki-18-corpus — Wikipedia 2018 dump in 100-word chunks
• Retrieval index: PeterJinGo/wiki-18-e5-index — e5-base-v2 dense index over the same corpus
• Source QA: RUC-NLPIR/FlashRAG_datasets (HotpotQA train, MuSiQue, 2WikiMultiHopQA)
• HotpotQA dev (distractor variant): hotpotqa/hotpot_qa

License

Inherits source dataset licenses:

• HotpotQA, wiki-18-corpus: CC BY-SA 4.0
• MuSiQue: CC BY 4.0
• 2WikiMultiHopQA: Apache 2.0
• FlashRAG aggregate: see source repo

If you use this dataset, please cite the original works for HotpotQA, MuSiQue, 2WikiMultiHopQA, and FlashRAG, plus PeterJinGo/wiki-18-corpus for the retrieval corpus.