Iman998/XhotpotQA-GLM52-Judge-V1

XHotpotQA · GLM-5.2 Judge

AUDIT SNAPSHOT · V1

Sanitized, language-balanced translation-quality audit

2,760 scored units · stable hash joins · public reasoning removed

AUDIT · V1 ROWS · 2,760 SCORED · 100% MODEL · requested alias only COMPARISON · independent / unpaired

Navigate: Overview · Load · Methodology · Results · Schema · Provenance · Limitations · Citation · Paper · Collection

Quality-label boundary. This dataset contains LLM-generated audit annotations, not QA-system predictions and not human gold labels. The endpoint was requested with model string glm-5.2; the provider-resolved model revision was not recorded.

Dataset at a glance

This release contains the complete V1 translation audit used in the XHotpotQA resource analysis. Each row evaluates one translated paragraph, question, or short answer against its English source on a 0–100 scale.

The public payload removes hidden reasoning and raw texts. Reproducible joins use stable XHotpotQA identifiers plus exact UTF-8 SHA-256 hashes of the source and candidate strings.

This V1 audit and the separately released V2 audit use independently drawn, balanced samples. The audits are independent and unpaired; their mean difference must not be interpreted as an instance-level or causal V2 improvement estimate.

The public annotation payload is frozen at Hub revision ba891ae62ed989606c9fc2fd5f08f9e88ef37547.

AUDIT UNITS 2,760
complete public snapshot

TARGET LANGUAGES 23
balanced sampling

SCORED 100%
2,760 of 2,760

WEIGHTED MEAN 92.097
0–100 judge scale

Quickstart

from datasets import load_dataset

AUDIT_REVISION = "ba891ae62ed989606c9fc2fd5f08f9e88ef37547"

audit = load_dataset(
    "Iman998/XhotpotQA-GLM52-Judge-V1",
    "xhotpotqa_glm52_judge_v1",
    split="audit",
    revision=AUDIT_REVISION,
)

print(audit.num_rows)  # 2760
print(audit.features)
print(audit.filter(lambda row: row["unit"] == "paragraph")[0])

Join rows to the default V1.1 configuration of the pinned XHotpotQA snapshot using instance_id, unit, and paragraph_id. V1.1 retains the V1 translations and adds original paragraph source_sentences. Verify exact joined strings with source_text_sha256 and candidate_text_sha256.

Methodology and sampling design

Sampling was deterministic and balanced over 23 target languages.

Unit	Per language	Total
paragraph	80	1,840
question	20	460
answer	20	460
all units	120	2,760

• Sampling seed: 20260810
• Train-derived units: 1,866
• Validation-derived units: 894
• Successfully scored: 2,760
• Failed final rows: 0

Paragraph source text was recovered from the pinned English HotpotQA context by source ID and title. Question and answer sources came from the original English fields.

Results

Unit	Count	Mean	Median	SD	Below 60	Below 80	At least 90
paragraph	1,840	90.913	94	9.47	24	182	1,358
question	460	93.459	96	10.71	12	35	387
answer	460	95.474	100	14.47	21	28	416
weighted overall	2,760	92.097	—	—	—	—	—

The overall mean reflects the audit allocation: two-thirds paragraphs, one-sixth questions, and one-sixth answers. Unit means are the primary quantities.

Per-language means

Language	Paragraph	Question	Answer	Overall
Arabic	88.05	94.35	99.60	91.03
Bengali	89.11	91.95	99.50	91.32
German	93.33	93.95	89.50	92.79
Greek	88.30	96.90	94.75	90.81
Spanish	95.08	96.05	96.00	95.39
Persian	87.99	97.90	95.75	90.93
French	93.95	96.35	97.50	94.94
Hindi	90.84	90.10	98.75	92.03
Indonesian	93.12	92.55	91.50	92.76
Italian	94.53	94.95	98.80	95.31
Japanese	89.61	94.30	95.25	91.33
Korean	91.42	93.40	98.75	92.97
Dutch	91.92	94.20	95.50	92.90
Polish	91.35	93.00	90.75	91.53
Portuguese	95.44	98.45	95.00	95.87
Russian	86.36	94.05	98.35	89.64
Swedish	89.76	91.60	88.75	89.90
Swahili	87.78	90.95	94.25	89.38
Thai	90.94	88.70	96.00	91.41
Turkish	88.31	87.85	94.00	89.18
Urdu	91.04	92.25	97.50	92.32
Vietnamese	92.46	94.50	91.25	92.60
Mandarin Chinese	90.30	91.25	98.90	91.89

Machine-readable full-precision results are included in tables/SUMMARY.json and tables/BY_LANGUAGE.csv. tables/BOOTSTRAP.json contains the deterministic within-language-by-unit percentile bootstrap.

Public schema

Field	Description
judge_record_id	Deterministic hash of version and unit identity
dataset_version, dataset_revision	Source dataset locator and candidate revision
instance_id, source_id, source_split	Stable dataset identifiers
target_language, unit, paragraph_id	Audit stratum and unit identity
score	Integer source-referenced score from 0 to 100
score_origin	Where the historical parser found the score
judge_explanation	Visible short explanation, if returned
source_text_sha256, candidate_text_sha256	Exact join-verification hashes
requested_judge_model	Requested API alias
resolved_judge_revision	Null because the endpoint revision was not persisted
judge_prompt_version, judge_prompt_sha256	Exact scoring contract identifiers
raw_artifact_sha256, raw_line_number	Private-archive lineage without exposing raw content

The release does not contain source_text, candidate_text, hidden reasoning, endpoint URL, request errors, API keys, or raw logs.

Shape-only annotation preview

Angle-bracketed values show the joinable public structure without reproducing source or candidate text.

{
  "judge_record_id": "<deterministic record hash>",
  "dataset_version": "v1",
  "instance_id": "<XHotpotQA instance identifier>",
  "target_language": "<language code>",
  "unit": "paragraph | question | answer",
  "paragraph_id": "<paragraph identifier or null>",
  "score": 0,
  "score_origin": "<parser provenance>",
  "judge_explanation": "<visible explanation if returned>",
  "source_text_sha256": "<SHA-256>",
  "candidate_text_sha256": "<SHA-256>",
  "requested_judge_model": "glm-5.2",
  "resolved_judge_revision": null
}

Score provenance

Origin	Rows
visible explicit SCORE: line	2,746
explicit score in reasoning fallback	4
last-integer reasoning fallback	10

Fourteen rows have no visible explanation and are explicitly labeled as reasoning-derived. No reasoning text is released.

Model provenance

Setting	Value
Requested model string	glm-5.2
Provider-resolved revision	not recorded
Temperature	0.0
Seed	20260810
Maximum output tokens	4,000
Concurrent workers	1
Final completion	2,760 of 2,760

The official GLM-5.2 model card documents the public model family, but it does not prove which checkpoint a third-party alias resolved to. The release therefore uses requested_alias_only_provider_revision_not_recorded as its identity status.

Historical prompt contract

The historical implementation produces one source-referenced score. An early code comment describing two scores is stale and is not supported by these records.

Paragraph and question system prompt

You are a meticulous bilingual translation judge. Your task is to assign a single integer score from 0 to 100 (0 = unusable, 100 = excellent) for a candidate translation, using only the source text and the target candidate.
Scoring rubric (decide internally; do not print sub-scores):
Adequacy/Faithfulness (60%): Candidate conveys all meaning from the source; no omissions/additions; no contradictions. Use your bilingual understanding to compare source meaning with the target text.
Terminology, Entities, Numbers (15%): Names, numbers, dates, units, and placeholders preserved and correct.
Fluency/Grammar (20%): Natural, grammatical writing in the target language.
Style/Register (5%): Tone/register appropriate to the source intent and genre.
Critical error floors:
Wrong language or untranslated copy of source -> <=10
Direct contradiction/mistranslation of key meaning -> <=40
Hallucinated content not grounded in source -> <=40
Loss/corruption of critical numbers, names, or placeholders -> <=60
Guidance: Favor semantic fidelity to the source over cleverness. Do not invent facts. Minor orthographic or punctuation quirks are minor unless they change meaning.
Round to the nearest integer and clamp to [0,100]. If inputs are empty or unusable, return 0.

Output format: respond with a short explanation (1-3 sentences) of the strengths and weaknesses of the candidate translation, then on a new final line output exactly "SCORE: <integer>" where <integer> is the score from 0 to 100.

SHA-256: e939a00a58e2347b60e9dc04ad7278b88430d41e4a2aea8749023c08631db5af

Short-answer system prompt

You are a meticulous bilingual translation judge for short answers. You are given an English question ("context_question") and its short English answer ("text") alongside a translated answer ("model_output"). Your task is to judge how well "model_output" translates the English answer into the target language.
Keep in mind: the answer is a concise factual response (a name, date, number, place, or yes/no). The translation must preserve the same factual content as the English answer, in the target language, while remaining natural.
Scoring rubric (decide internally; do not print sub-scores):
Faithfulness (70%): The translated answer conveys the same fact(s) as the English answer. For yes/no answers the polarity must match.
Terminology, Entities, Numbers (20%): Names, numbers, dates, and entities preserved and correct.
Fluency (10%): Natural, grammatical form in the target language for a short answer.
Critical error floors:
Wrong language or untranslated English copy -> <=10
Opposite yes/no polarity or completely different fact -> <=20
Hallucinated content not grounded in the English answer -> <=40
Loss/corruption of critical numbers, names, or dates -> <=60
Guidance: Names that are conventionally kept in Latin script in the target language should not be penalized. Minor transliteration differences are acceptable.
Round to the nearest integer and clamp to [0,100]. If inputs are empty or unusable, return 0.

Output format: respond with a short explanation (1-3 sentences) of the strengths and weaknesses of the candidate answer translation, then on a new final line output exactly "SCORE: <integer>" where <integer> is the score from 0 to 100.

SHA-256: 7e560af3c545c22fee93f78f3b621326e04c92ddcd61844c79ee45994320a525

The phrase “critical error floors” is preserved literally for reproducibility. Because each rule uses an upper-bound sign, the criteria operate as score ceilings.

User payload

{
  "source_language": "en",
  "target_language": "<language code>",
  "text": "<English source>",
  "model_output": "<translated candidate>"
}

Answer requests additionally contain context_question with the English question.

Limitations

• One LLM judge is not a substitute for bilingual human annotation or inter-annotator agreement.
• The provider-resolved model revision and server configuration are unavailable.
• Temperature zero and a seed do not guarantee deterministic third-party endpoint behavior.
• The score parser historically allowed an integer fallback; score_origin makes those cases visible.
• The rubric can apply transliteration expectations unevenly across languages and entity types.
• The sample is balanced but not a census of every translation.
• The V1 source Parquet hashes were not written into the original judge run; the public revision is a retrospective candidate locator and is labeled accordingly.

License

These annotations derive from CC BY-SA 4.0 HotpotQA/XHotpotQA content and are released under CC BY-SA 4.0. The accompanying builder software is MIT-licensed.

Citation

Use the XHotpotQA and HotpotQA BibTeX entries in the canonical V1.1 dataset card. The XHotpotQA paper is archived as arXiv:2608.27481, DOI 10.48550/arXiv.2608.27481. In the audit-artifact description, report Iman998/XhotpotQA-GLM52-Judge-V1, configuration xhotpotqa_glm52_judge_v1, frozen revision ba891ae62ed989606c9fc2fd5f08f9e88ef37547, the requested glm-5.2 alias, and the independent/unpaired sampling design.

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
Incomplete release candidate
Open card →

Judge · V1
Current independent audit
Open card →

Judge · V2
Independent comparison audit
Open card →

Resources

• Frozen judge snapshot: https://huggingface.co/datasets/Iman998/XhotpotQA-GLM52-Judge-V1/tree/ba891ae62ed989606c9fc2fd5f08f9e88ef37547
• Pinned XHotpotQA V1.1 source: https://huggingface.co/datasets/Iman998/XhotpotQA/tree/1d29e7918cf1acc045726c70fddba82371833090
• Code: https://github.com/Iman998/XhotpotQA
• Paper: https://arxiv.org/abs/2608.27481
• Paper DOI: https://doi.org/10.48550/arXiv.2608.27481
• Official GLM-5.2 announcement: https://z.ai/blog/glm-5.2
• Official GLM-5.2 model card: https://huggingface.co/zai-org/GLM-5.2