Iman998/XhotpotQA-GLM52-Judge-V2

XHotpotQA · GLM-5.2 Judge

AUDIT SNAPSHOT · V2

Sanitized audit of the Gemma 4 31B-generated release candidate

Clean 2,760-unit set · malformed partial-answer run excluded

AUDIT · V2 RC1 ROWS · 2,760 SCORED · 100% MODEL · requested alias only EXCLUDED · 240 malformed rows COMPARISON · independent / unpaired

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

Quality-label and comparison boundary. This is an audit of V2 RC1, not a human-certified quality label. The requested API model string was glm-5.2; the provider-resolved checkpoint revision was not stored. V1 and V2 audit samples are independent and must not be described as paired.

Dataset at a glance

This release provides the clean final 2,760-row translation audit for XHotpotQA V2 RC1. Paragraphs, questions, and short answers are scored against their English source on a 0–100 scale.

The release builder reconstructs the valid final set from locked artifacts, excludes the known malformed partial answer run, removes hidden reasoning and raw text, and records a cryptographic locator for every source/candidate pair.

The public annotation payload is frozen at Hub revision 0f9cd568fabd7f7ad3b3d9a72e31ae8aeb936840.

AUDIT UNITS 2,760
clean final set

TARGET LANGUAGES 23
balanced sampling

EXCLUDED 240
malformed partial-answer rows

WEIGHTED MEAN 94.677
0–100 judge scale

Quickstart

from datasets import load_dataset

AUDIT_REVISION = "0f9cd568fabd7f7ad3b3d9a72e31ae8aeb936840"

audit = load_dataset(
    "Iman998/XhotpotQA-GLM52-Judge-V2",
    "xhotpotqa_glm52_judge_v2",
    split="audit",
    revision=AUDIT_REVISION,
)

print(audit.num_rows)  # 2760

thai_answers = audit.filter(lambda row: row["target_language"] == "th" and row["unit"] == "answer")
print(thai_answers["score"])

Join against the pinned XHotpotQA V2 RC1 snapshot using instance_id, unit, and paragraph_id. Validate the joined strings with their exact SHA-256 fields.

Methodology and sampling design

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

• Target languages: 23
• Sampling seed: 20260810
• Train-derived units: 1,883
• Validation-derived units: 877
• Successfully scored: 2,760
• Failed final rows: 0

Artifact recovery

The raw judge_all.records.jsonl is not a publication artifact. It contains:

• the valid 1,840 paragraph rows;
• the valid 460 question rows;
• 239 scored partial answer rows produced by an incorrect source-unit path, plus one failed row from the same partial run.

The 239 scored rows compare translated answers against the English question instead of the English answer and have a misleading mean of 4.736; the remaining row has no score because its request failed. The clean release discards all 240 rows and uses the dedicated 460-row answer artifact produced with the English answer plus question context.

During paragraph/question evaluation, 276 transient API failures remained after the first attempt sequence. The retry run rescored all 276 successfully. Final paragraph and question coverage is therefore complete.

Results

Unit	Count	Mean	Median	SD	Below 60	Below 80	At least 90
paragraph	1,840	94.468	96	7.53	14	64	1,679
question	460	96.009	98	8.12	6	19	421
answer	460	94.183	100	19.46	29	31	419
weighted overall	2,760	94.677	—	—	—	—	—

Paragraph and question distributions are concentrated near the top of the rubric. Answers have a high median but a wider lower tail, so the answer mean should not be summarized as uniformly high quality across languages.

Per-language means

Language	Paragraph	Question	Answer	Overall
Arabic	95.03	97.45	96.75	95.72
Bengali	95.15	95.15	95.40	95.19
German	95.61	98.30	99.75	96.75
Greek	84.72	89.60	86.60	85.85
Spanish	95.42	95.30	94.10	95.18
Persian	94.29	94.95	91.75	93.97
French	96.51	96.90	100.00	97.16
Hindi	93.95	97.00	99.25	95.34
Indonesian	96.51	97.40	98.50	96.99
Italian	95.89	95.70	96.75	96.00
Japanese	95.36	94.80	99.75	96.00
Korean	96.33	97.15	99.25	96.95
Dutch	94.14	96.45	92.50	94.25
Polish	94.58	99.20	98.50	96.00
Portuguese	96.25	97.45	90.50	95.49
Russian	93.85	96.35	100.00	95.29
Swedish	92.65	96.50	98.40	94.25
Swahili	93.34	92.60	94.65	93.43
Thai	94.47	96.35	65.30	89.92
Turkish	94.15	98.25	98.15	95.50
Urdu	93.33	95.80	90.10	93.20
Vietnamese	96.54	95.75	95.10	96.17
Mandarin Chinese	94.70	93.80	85.15	92.96

Machine-readable full-precision values are in tables/SUMMARY.json and tables/BY_LANGUAGE.csv. Confidence intervals are in tables/BOOTSTRAP.json.

Interpreting the language tails

The visible judge explanations support three restrained observations:

• Several low Greek paragraph scores identify mixed-script intrusions, untranslated fragments, or meaning reversals.
• Thai answer scores are strongly affected by unchanged Latin-script names and the judge's expectation of Thai transliteration.
• Some Chinese answer penalties similarly concern conventional entity or title rendering.

V2 contains 179 exact English-source copies among the 460 sampled answers. Many are legitimate proper names; others can reflect the historical source-copy fallback. An exact copy is therefore a review signal, not automatically an error. The judge itself may apply transliteration conventions inconsistently.

Descriptive V1 comparison

Unit	V1 mean	V2 mean	V2 minus V1
paragraph	90.913	94.468	+3.555
question	93.459	96.009	+2.550
answer	95.474	94.183	−1.291
weighted overall	92.097	94.677	+2.580

This comparison is not paired. Overlap on source ID, target language, and unit is only nine paragraph units, zero questions, and one answer. The values describe two balanced samples; they do not identify a causal Gemma 4 improvement. The answer audit in particular does not improve uniformly.

Public schema

Field	Description
judge_record_id	Deterministic hash of version and unit identity
dataset_version, dataset_revision	V2 RC1 and locked-input fingerprint
instance_id, source_id, source_split	Stable source locators
target_language, unit, paragraph_id	Audit identity and stratum
score, score_origin	Integer rating and parser provenance
judge_explanation	Visible explanation only
source_text_sha256, candidate_text_sha256	Exact text-join verification
requested_judge_model	Requested API string glm-5.2
resolved_judge_revision	Null; not persisted by the endpoint run
judge_prompt_version, judge_prompt_sha256	Exact historical scoring contract
run_group	Paragraph/question recovery or dedicated answer run
raw_artifact_sha256, raw_line_number	Private-archive lineage

The dataset omits source/candidate text duplication, hidden reasoning, raw endpoint details, errors, credentials, and 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": "v2",
  "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,
  "run_group": "<recovery or dedicated answer run>"
}

Score provenance

Origin	Rows
visible explicit SCORE: line	2,756
last-integer reasoning fallback	4

The four reasoning-fallback rows have no visible explanation. Their origin is retained, but the reasoning text is not released.

Model provenance

Setting	Value
Requested judge model	glm-5.2
Provider-resolved revision	not recorded
Temperature	0.0
Seed	20260810
Maximum output tokens	4,000
Concurrent workers	1

The official GLM-5.2 sources establish the public model family, not the identity behind a third-party alias. This release therefore makes no immutable-checkpoint claim.

Historical prompt contract

The artifacts contain one source-referenced score. They do not contain a second reference-free score.

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 historical phrase “critical error floors” is reproduced exactly. Because the rules specify upper bounds, they function as score ceilings.

User payload

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

The answer request additionally carries the original English question as context_question.

Limitations

• Scores come from one uncalibrated LLM judge rather than bilingual human annotators.
• The provider-resolved judge model and server revision are unknown.
• The V1/V2 samples are almost entirely unpaired.
• Language-specific entity and transliteration preferences can affect scores.
• High medians coexist with meaningful lower-tail failures.
• The audit covers a balanced sample, not all 22,836 RC1 rows.
• The underlying V2 dataset remains incomplete and status-bearing.

License

These annotations derive from CC BY-SA 4.0 HotpotQA/XHotpotQA content and are released under CC BY-SA 4.0. Builder code 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-V2, configuration xhotpotqa_glm52_judge_v2, frozen revision 0f9cd568fabd7f7ad3b3d9a72e31ae8aeb936840, the requested glm-5.2 alias, the independent/unpaired sampling design, and the exclusion of the 240 malformed partial-answer rows.

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 source dataset
Open card →

Judge · V1
Independent comparison audit
Open card →

Judge · V2
Current independent audit
Open card →

Resources

• Frozen judge snapshot: https://huggingface.co/datasets/Iman998/XhotpotQA-GLM52-Judge-V2/tree/0f9cd568fabd7f7ad3b3d9a72e31ae8aeb936840
• Pinned V2 RC1 source: https://huggingface.co/datasets/Iman998/XhotpotQA-V2/tree/b05ba394ad7312e85625624c90d10258cbab31af
• 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
• Gemma 4 technical report: https://arxiv.org/abs/2607.02770