{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2607.29677", "kind": "paper", "url": "https://arxiv.org/abs/2607.29677"}, {"id": "gh:run-llama/extractbench", "kind": "repo", "url": "https://github.com/run-llama/ExtractBench"}, {"id": "hf:llamaindex/extractbench", "kind": "dataset", "url": "https://huggingface.co/datasets/llamaindex/ExtractBench"}], "authors": ["Boyang Zhang", "Adrian Lyjak", "Eli Stewart", "Zhaoqi Li", "Simon Suo"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2607.29677", "source_url": "https://arxiv.org/abs/2607.29677"}, "categories": ["Factuality", "General AI", "Language & Knowledge", "cs.AI"], "description": {"en": "ExtractBench evaluates schema-guided extraction from enterprise documents. Given a document and a user-defined JSON schema, systems must return schema-valid JSON with correct values, include every record of repeated structures, mark missing fields as null, and provide source evidence. The benchmark includes 4,869 pages across 370 documents, 8 business domains, and 67 document types, with tags for challenge, perception, table structure, length, and domain. Scoring uses unified value F1 for value accuracy and word- and page-level F1 for grounding."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.29677", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.29677", "title": "ExtractBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2607.29677", "modality": null, "name": "ExtractBench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "High", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "unknown", "origin_record_id": "bm_extractbench_2ecdf546", "origin_source": "arxiv", "recognition_confidence": "0.95", "record_sha256": "43945a741dc2f0362e3b9cd559d0fb1574ccb6f58845d64e72fbdeb0f802bdf7", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2607.29677"}, "publisher": null, "released": "2026-07-31", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.29677", "source_url": "https://arxiv.org/abs/2607.29677"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-29677", "source": "claire_radar", "source_benchmark_id": "2607.29677", "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": "https://github.com/run-llama/ExtractBench", "githubScope": "benchmark_repo", "githubStars": 105, "hfDailySubmittedAt": "2026-08-03T00:00:00.000Z", "hfDataset": "llamaindex/ExtractBench", "hfDatasetDownloads": 16491, "hfDatasetLikes": 33, "hfPaperUpvotes": 25, "hfPaperUrl": "https://huggingface.co/papers/2607.29677", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "hfDatasetDownloads": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "hfDatasetLikes": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "hfPaperUpvotes": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}}, "status": "fresh"}, "capabilities": ["Factuality"], "confidence": "High", "construction": "Original Synthetic", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "89d3015ed6801258917c496c948fb9c04f09e17fbe7b0257ca0d8f949129f8fe", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "ExtractBench evaluates schema-guided extraction from enterprise documents. Given a document and a user-defined JSON schema, systems must return schema-valid JSON with correct values, include every record of repeated structures, mark missing fields as null, and provide source evidence. The benchmark includes 4,869 pages across 370 documents, 8 business domains, and 67 document types, with tags for challenge, perception, table structure, length, and domain. Scoring uses unified value F1 for value accuracy and word- and page-level F1 for grounding.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "public artifact URL"], "snippet": "We present ExtractBench, a benchmark for schema-guided extraction and, to our knowledge, the first to score value accuracy, record completeness at scale, grounding, and measured cost together."}, "familyId": "bmf_b7c261db99d3", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [{"evidence": [{"basis": "existing-code-link", "linkedUrl": "https://github.com/run-llama/ExtractBench", "url": "https://github.com/run-llama/ExtractBench"}, {"basis": "existing-metric-association", "linkedUrl": "https://github.com/run-llama/ExtractBench", "url": "https://huggingface.co/papers/2607.29677"}, {"basis": "linked-from-report", "linkedUrl": "https://github.com/run-llama/ExtractBench", "url": "https://arxiv.org/abs/2607.29677"}], "scope": "benchmark_repo", "url": "https://github.com/run-llama/ExtractBench", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_extractbench_2ecdf546", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/run-llama/ExtractBench", "data": "https://huggingface.co/datasets/llamaindex/ExtractBench", "hfPaper": "https://huggingface.co/papers/2607.29677", "paper": "https://arxiv.org/abs/2607.29677", "pdf": "https://arxiv.org/pdf/2607.29677", "project": null, "report": "https://arxiv.org/abs/2607.29677"}, "metrics": [], "motivation": "Enterprise workflows increasingly rely on agents for \\emph{schema-guided extraction}: given a document and a user-defined schema, the agent faithfully follows the schema to produce the correct output with source evidence as grounding metadata.", "name": "ExtractBench", "oneLine": "ExtractBench evaluates schema-guided extraction from enterprise documents. Given a document and a user-defined JSON schema, systems must return schema-valid JSON with correct values, include every record of repeated structures, mark missing fields as null, and provide source evidence. The benchmark includes 4,869 pages across 370 documents, 8 business domains, and 67 document types, with tags for challenge, perception, table structure, length, and domain. Scoring uses unified value F1 for value accuracy and word- and page-level F1 for grounding.", "paperTitle": "ExtractBench: A Benchmark for Schema-Guided Enterprise Document Extraction", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2607.29677", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "LlamaIndex", "organizationType": "company-research-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/run-llama/ExtractBench"}], "ranking": {"90d": {"confidence": "High", "coverage": 1.0, "datasetDownloadRank": 1, "datasetRankPopulation": 85, "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": 0.6224852885180543, "value": 105}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": 1.0, "value": 16491}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.57568809544959, "value": 25}}, "confidence": "High", "coverage": 1.0, "rank": 6, "score": 73}, "method": "attention-ranking-v13", "rank": 6, "score": 73}}, "readiness": "Runnable", "recognitionConfidence": 0.95, "relation": "introduces", "releasedAt": "2026-07-31", "source": {"authors": ["Boyang Zhang", "Adrian Lyjak", "Eli Stewart", "Zhaoqi Li", "Simon Suo"], "categories": ["cs.AI"], "id": "2607.29677", "title": "ExtractBench: A Benchmark for Schema-Guided Enterprise Document Extraction", "type": "arxiv", "url": "https://arxiv.org/abs/2607.29677"}, "sourceUpdatedAt": "2026-08-05T17:48:49Z", "topics": ["cs.AI"], "venueAttempts": [], "whyItMatters": "Enterprise workflows increasingly rely on agents for schema-guided extraction, where errors can lead to wrong payments or decisions. ExtractBench addresses the lack of a benchmark that jointly measures value accuracy, completeness, grounding, and cost, providing a standardized evaluation for comparing extraction systems on realistic document types and lengths."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
