Iman998/XhotpotQA-V2

XHotpotQA V2

AUDITED RC1 · INCOMPLETE

Cross-lingual multi-hop QA over mixed-language evidence

Gemma 4 31B · source-aligned fields · transparent release gate

STATUS · RC1 ROWS · 22,836 GENERATOR · Gemma 4 31B MISSING · 230 LICENSE · CC BY-SA 4.0

Navigate: Overview · Coverage · Load · Schema · Methodology · Quality · Citation · Collection

RC1 release warning. This is a transparent release candidate, not the corrected canonical V2. The locked files contain 22,836 of 23,066 expected HotpotQA sources. No missing row is generated, copied, or silently hidden by the release builder.

Dataset at a glance

XHotpotQA V2 is a translation-derived benchmark for controlled cross-lingual multi-hop QA. A question and answer use one assigned language while candidate paragraphs may use different languages inside the same instance. Stable paragraph and sentence identifiers preserve the HotpotQA supporting chain.

RC1 adds the English source material needed for direct auditing:

• source_question and source_answer on every released row;
• source_title and source_sentences for every candidate paragraph;
• row-level structural and quality flags;
• source, input, and release checksums;
• a complete manifest for every expected-but-absent source.

RELEASED 22,836
audited RC1 rows

EXPECTED 23,066
HotpotQA sources

MISSING 230
explicitly manifested

COVERAGE 99.003%
not canonical-complete

Coverage

Split	Released	Expected	Missing	Coverage
train (HotpotQA hard)	15,433	15,661	228	98.544%
validation (distractor)	7,403	7,405	2	99.973%
total	22,836	23,066	230	99.003%

Both missing validation sources are inherited HotpotQA anomalies: one contains a blank sentence inside a supplied distractor paragraph, and one contains supporting-fact index 902 outside the annotated paragraph. Across both splits, 50 missing rows are triggered by a blank source sentence or an out-of-bounds source support annotation; the other 180 are clean-source training omissions from the incomplete generation run. MISSING_SOURCE_MANIFEST.json contains the exact absent source IDs. The 50/180 attribution reported here was recomputed by joining those IDs to the official HotpotQA rows and revalidating their sentence arrays and supporting-fact indices; the original run ledger did not encode every retrospective reason class.

Quickstart

from datasets import load_dataset

DATA_REVISION = "b05ba394ad7312e85625624c90d10258cbab31af"

dataset = load_dataset(
    "Iman998/XhotpotQA-V2",
    "xhotpotqa_v2_audited_rc1",
    revision=DATA_REVISION,
)

row = dataset["validation"][0]
print(row["question"], row["question_language"])
print(row["source_question"])

for paragraph in row["candidates"]:
    print(paragraph["language"], paragraph["source_title"])
    print(paragraph["source_sentences"][0])
    print(paragraph["sentences"][0])

The example pins the published RC1 snapshot b05ba394ad7312e85625624c90d10258cbab31af. Record this revision together with the release-manifest fingerprint in every experiment instead of loading a moving main branch.

Record structure

Field	Meaning
id	Generated XHotpotQA instance identifier
source_id, source_split, source_position	Stable location in the pinned HotpotQA source
question, answer	Translated QA pair
question_language, answer_language	ISO-like language codes used by the release
source_question, source_answer	Original English HotpotQA text
candidates	Ordered translated paragraphs plus original titles and sentences
supporting_facts	Stable paragraph/sentence links with source titles and bounds checks
status	accepted, review_required, or quarantined
structural_flags	Deterministic structural/source-alignment findings
quality_flags	Deterministic review signals such as source-copy output
source_record_sha256	Checksum of the canonical source record
input_record_checksum_sha256	Generator-provided semantic checksum
release_record_sha256	Checksum of the normalized public row

Candidate structure

Each candidate contains:

paragraph_id, candidate_index,
source_title, source_sentences,
title, sentences,
language, language_name,
source_match

The release builder takes source_title and source_sentences from the pinned HotpotQA source. It separately verifies the values recorded by generation and sets source_match; a mismatch is never silently accepted.

Shape-only record preview

Angle-bracketed values below describe the published structure; they are not a substitute for loading a released row.

{
  "id": "<generated instance identifier>",
  "source_id": "<HotpotQA source identifier>",
  "question": "<translated question>",
  "answer": "<translated answer>",
  "question_language": "<assigned language code>",
  "source_question": "<original English question>",
  "candidates": [
    {
      "paragraph_id": "p00",
      "source_title": "<English source title>",
      "source_sentences": ["<English source sentence>"],
      "sentences": ["<translated sentence>"],
      "language": "<assigned language code>"
    }
  ],
  "supporting_facts": ["<stable paragraph/sentence links>"],
  "status": "accepted | review_required | quarantined"
}

Status and quality policy

accepted means the row passed the deterministic checks and has no content-level review flag. review_required means structure remains usable but an automatic quality signal needs inspection. quarantined means at least one structural or source-alignment condition failed.

ACCEPTED
No blocking flag and no content-level review flag.

REVIEW REQUIRED
Structurally usable; automatic quality signal requires inspection.

QUARANTINED
Structural or source-alignment condition failed.

Representative structural flags include:

• xhotpot:input_checksum_mismatch
• xhotpot:candidate_count_mismatch
• xhotpot:source_sentences_mismatch
• xhotpot:sentence_count_mismatch
• xhotpot:support_annotation_mismatch
• source:blank_source_sentence
• source:support_index_out_of_range

Representative quality flags include:

• xhotpot:question_source_copy
• xhotpot:answer_source_copy
• xhotpot:paragraph_sentence_source_copy
• provenance:assignment_manifest_hash_missing

Flags prefixed with source: describe an inherited source condition. Flags prefixed with xhotpot: describe a generated-record or transformation condition. This distinction prevents a HotpotQA anomaly from being misreported as a translation failure.

Methodology and generation provenance

The supplied run identifies the generator as Gemma 4 31B Instruct, served through vLLM's OpenAI-compatible API.

Property	Recorded value
Run configuration model	google/gemma-4-31B-it
Served model ID stored in rows	gemma-4-31B-it
Operator-recorded revision	gemma-4-31B-it-vllm-v0.19.1
Prompt version	xhotpotqa-translation-v2.0
Prompt SHA-256	623496d198d7850c244ff4e2303b7ba9b61548499ce10256ae6691a6b58e71f3
Seed	20260810
Thinking output	disabled in recorded chat-template options
Recorded generation interval	2026-08-12T20:55:00Z to 2026-08-13T21:23:54Z

The locked rows contain 22,379 records produced at temperature 0.0, 354 retry records at 0.2, and 103 retry records at 0.3. The operator report describes vLLM 0.19.1 with tensor parallelism over two GPUs. No immutable Hub model commit was persisted, so the served revision string is provenance, not a cryptographic checkpoint identity.

Literal V2 translation system prompt

You are the deterministic translation component of a multilingual QA dataset. Preserve named entities, numbers, dates, yes/no polarity, and sentence boundaries. Do not answer the question and do not add explanations. The user request contains a response_schema; return exactly one valid JSON object that satisfies it, with no additional keys and no Markdown.

Single strings must return exactly one translation string. Paragraph requests must return exactly one translations array with the same cardinality as the source sentence array.

Locked source and release integrity

Input	SHA-256
HotpotQA train v1.1	26650cf50234ef5fb2e664ed70bbecdfd87815e6bffc257e068efea5cf7cd316
HotpotQA distractor validation	e3da074df24e8369009918aa5cdbdd254dadcde4c63f7569d36afd6f2268caa8
V2 train JSONL	dd1d5bb5950cfe3ca5d013685f9d6e71d1059bde0e5a316462e26a546d491270
V2 validation JSONL	86542d9918dab1e0587683b51dfa7091a6e8b77171283c66caca35ed70ac931a

The builder writes Parquet into a private staging directory, validates counts and identifiers, hashes every output, and only then atomically installs the completed release directory. Existing output directories are never overwritten.

Quality audit

A separate GLM-5.2 judge release contains a language-balanced audit of 1,840 paragraph, 460 question, and 460 answer translations. V2 mean scores are 94.468, 96.009, and 94.183, respectively. The V1 and V2 samples are almost entirely different source/language assignments, so this is an independent descriptive audit—not a paired improvement estimate.

See XHotpotQA-GLM52-Judge-V2 for the sanitized scores, prompt hashes, and limitations.

Release status

RC1 deliberately fails the canonical-completeness gate:

• 230 expected sources are absent;
• every recorded assignment-manifest hash is empty;
• the historical concurrent retry_count field is not a reliable per-record count;
• structural-passed in the generator means structural validation only, not semantic adequacy or target-language compliance;
• the historical fallback could retain source English when translation failed.

Use RC1 for auditing, code validation, and explicitly status-aware experiments. Do not describe it as the complete or corrected canonical V2.

Intended uses

• Mixed-language multi-hop QA research.
• Evidence-language and script-robustness analysis.
• Reader and selector diagnostics with stable supporting facts.
• Translation-quality and provenance research using source-aligned fields.

Out-of-scope uses

• Treating automatic scores as human ground truth.
• Native-language cultural or information-seeking claims.
• Safety-critical decisions.
• Ignoring row status or the missing-source manifest.

Limitations

• The content originates from English Wikipedia through HotpotQA; translations do not create native information needs.
• Translation and transliteration quality can differ by language, script, entity type, and answer type.
• Exact source copies can be correct for names or titles, so automatic source-copy flags require contextual review.
• The release is incomplete: clean-source omissions require recovery or regeneration, while source-triggered omissions require an explicit policy for malformed upstream rows.
• The accompanying LLM judge is a single uncalibrated model alias and may apply language-specific transliteration preferences inconsistently.

License and attribution

HotpotQA is distributed under CC BY-SA 4.0. XHotpotQA is a transformed, source-aligned resource and is distributed under the same license. Repository software is licensed separately under MIT.

Please cite HotpotQA and the XHotpotQA paper/repository. A persistent article/archive identifier will be added after deposit.

Citation

Use the XHotpotQA and HotpotQA BibTeX entries in the canonical V1.1 dataset card. In the artifact description, report Iman998/XhotpotQA-V2, configuration xhotpotqa_v2_audited_rc1, frozen revision b05ba394ad7312e85625624c90d10258cbab31af, and the incomplete RC1 status.

Release family

Browse V1.1, V2 RC1, and both independent GLM-5.2 audit snapshots in the XHotpotQA cross-lingual multi-hop QA collection.

V1.1
Canonical audited data
Open card →

V2 RC1
Current incomplete release
Open card →

Judge · V1
Independent GLM-5.2 audit
Open card →

Judge · V2
Independent GLM-5.2 audit
Open card →

Resources

• Frozen dataset snapshot: https://huggingface.co/datasets/Iman998/XhotpotQA-V2/tree/b05ba394ad7312e85625624c90d10258cbab31af
• Code: https://github.com/Iman998/XhotpotQA
• Gemma 4 technical report: https://arxiv.org/abs/2607.02770
• vLLM Gemma 4 guide: https://docs.vllm.ai/projects/recipes/en/stable/Google/Gemma4.html
• Paper/archive: forthcoming