README.md

metadata

license: apache-2.0
task_categories:
  - question-answering
language:
  - en
tags:
  - rag
  - cacheblend
  - multi-hop
  - kv-cache
size_categories:
  - n<1K

CacheBlend RAG Extended

Multi-hop QA splits derived from MuSiQue and 2WikiMQA (as bundled in the official CacheBlend repo), augmented with extra distractor chunks per query so that each example carries 20 context passages instead of the original 10.

Designed to amplify the contrast between full_reuse (no cross-chunk attention -> quality drop) and cacheblend (selective KV recompute -> quality recovers) while preserving the multi-hop questions and gold short answers.

Splits

split	n	chunks/query	~tokens/chunk	gold preserved vs source	structural preservation	CacheBlend ready
musique	150	20	598	100.00%	100%	PASS
wikimqa	200	20	560	103.81%	100%	PASS

Schema

Each row matches the bundled CacheBlend JSON format:

{
  "question": "Where was the author of Hannibal and Scipio educated at?",
  "ctxs": [
    {"title": "<wiki title>", "text": "<512-token passage>"},
    ...  // 20 entries
  ],
  "answers": ["Exeter College"]
}

How it was built

1. Start from musique_s.json (150 ex) and wikimqa_s.json (200 ex) — both already have the multi-hop questions and 10 gold/distractor passages per query that the CacheBlend paper used.
2. For each query, sample 10 additional chunks from OTHER queries in the same source (no chunk is reused twice within a query; titles deduped).
3. Shuffle the resulting 20 chunks so gold evidence is not always first.
4. Verify (per row): (i) chunks/query == 20, (ii) at least one gold-answer substring is preserved across the 20 chunks.

Reproduce with scripts/build_cacheblend_dataset.py --extra 10 in mkim0628/experiment-reproducer.

Suitability gates

The build script runs three CacheBlend-relevance checks. All splits must pass before upload:

• chunks_per_query_ge_10: at least 10 passages so KV reuse pays off
• gold_preservation_ratio_ge_0.95: ≥95 % of queries still have the gold-answer substring after distractor injection
• chunk_size_within_paper_range_300_800_tokens: chunk length matches the paper's 512-token spec

See suitability_report.json for the per-split numbers.