{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2609.11399", "kind": "paper", "url": "https://arxiv.org/abs/2609.11399"}], "authors": ["Shenbin Qian", "Yves Scherrer"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.11399", "source_url": "https://arxiv.org/abs/2609.11399"}, "categories": ["General AI", "Language & Knowledge", "cs.CL"], "description": {"en": "TransClean is a controlled benchmark of 9,900 noisy-and-clean translation output pairs for detecting and extracting clean translations from large language model outputs, including 8,800 synthetic and 1,100 manually curated instances."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.11399", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.11399", "title": "TransClean"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2609.11399", "modality": null, "name": "TransClean", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "false", "origin_record_id": "bm_transclean_110295c3", "origin_source": "arxiv", "recognition_confidence": "0.85", "record_sha256": "cb6e5f317ae5fba629d27c9ba5deb96c8b00c4870ab5f0f56a074e8df3555605", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-name-audit-deferred", "reviewed_at": "2026-09-14T00:43:53.625753Z", "source_url": "https://arxiv.org/abs/2609.11399"}, "publisher": null, "released": "2026-09-10", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.11399", "source_url": "https://arxiv.org/abs/2609.11399"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-11399", "source": "claire_radar", "source_benchmark_id": "2609.11399", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Machine Generated", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 0, "hfPaperUrl": "https://huggingface.co/papers/2609.11399", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "A translation-output cleaning benchmark is a niche post-processing topic, and no released data link or evaluator is supplied to support broader early visibility.", "score": 22}, "capabilities": [], "confidence": "Medium", "construction": "Original Synthetic", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-15T09:37:10.086794Z", "inputHash": "25f057e96e4070de71589c9ec1a9f3f921ad0bbd17f408afda073d6387c48b02", "model": "deepseek-v4-flash"}, "curation": {"decisionReason": "No dataset or code release is indicated, and the paper alone does not provide a reusable evaluation path.", "model": "deepseek-v4-pro", "policy": "source-evidence-v2", "reviewedAt": "2026-09-14T00:43:53.625753Z", "state": "ai-name-audit-deferred"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "TransClean is a controlled benchmark of 9,900 noisy-and-clean translation output pairs for detecting and extracting clean translations from large language model outputs, including 8,800 synthetic and 1,100 manually curated instances.", "displayEligible": false, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "evaluation protocol evidence"], "snippet": "Building on the observed patterns, we construct TransClean, a controlled benchmark of 9,900 pairs of noisy and clean translation outputs, comprising 8,800 synthetically generated instances and 1,100 manually curated authentic instances."}, "familyId": "bmf_b130209d5ada", "firstSeenAt": "2026-09-11", "heat": null, "id": "bm_transclean_110295c3", "indexedAt": "2026-09-11T00:48:15.671466Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2609.11399", "pdf": "https://arxiv.org/pdf/2609.11399", "project": null, "report": "https://arxiv.org/abs/2609.11399"}, "metrics": [], "motivation": "Large language models (LLMs) are increasingly used for machine translation, yet their outputs often contain additional text beyond the translation itself, such as language labels, explanations or bilingual repetitions, which we term translation noise.", "name": "TransClean", "oneLine": "TransClean is a controlled benchmark of 9,900 noisy-and-clean translation output pairs for detecting and extracting clean translations from large language model outputs, including 8,800 synthetic and 1,100 manually curated instances.", "paperTitle": "TransClean: A Benchmark for Detecting and Extracting Clean Translations from Large Language Model Outputs", "primaryDomain": "General AI", "publication": {"evidence": "Accepted to the Eleventh Conference on Machine Translation (WMT26)", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2609.11399", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "Eleventh Conference on Machine Translation (WMT26)", "verifiedAt": "2026-09-11T00:48:15.671466Z"}, "publications": [], "ranking": {"30d": {"confidence": "Low", "coverage": 0.3, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": 30, "historyDate": "2026-09-05", "rank": null, "score": null, "windowDays": 30}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": null, "value": null}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.0, "value": 0}}, "confidence": "Low", "coverage": 0.3, "rank": null, "score": 35}, "method": "attention-ranking-v13", "rank": null, "score": 35}, "90d": {"confidence": "Low", "coverage": 0.15, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": null, "value": null}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.0, "value": 0}}, "confidence": "Low", "coverage": 0.15, "rank": null, "score": 43}, "method": "attention-ranking-v13", "rank": null, "score": 43}}, "readiness": "Paper only", "recognitionConfidence": 0.85, "relation": "introduces", "releasedAt": "2026-09-10", "source": {"authors": ["Shenbin Qian and Yves Scherrer"], "categories": ["cs.CL"], "id": "2609.11399", "title": "TransClean: A Benchmark for Detecting and Extracting Clean Translations from Large Language Model Outputs", "type": "arxiv", "url": "https://arxiv.org/abs/2609.11399"}, "sourceUpdatedAt": "2026-09-10T11:32:18Z", "topics": ["cs.CL"], "venueAttempts": [{"decisionRaw": "Accepted to the Eleventh Conference on Machine Translation (WMT26)", "evidence": [{"level": "author-claim", "observedAt": "2026-09-11T00:48:15.671466Z", "rawValue": "Accepted to the Eleventh Conference on Machine Translation (WMT26)", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2609.11399"}], "reviewStatus": "accepted", "venueName": "Eleventh Conference on Machine Translation (WMT26)"}], "whyItMatters": "Model outputs often embed unwanted artifacts around usable translations, and separating clean translations from noise is a practical post-processing concern; a released paired corpus would let teams compare detection and extraction methods on common data."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
