{"record": {"aliases": ["DataComp-VLM: Improved Open Datasets for Vision-Language Models"], "artifacts": [{"id": "arxiv:2606.28551", "kind": "paper", "url": "https://arxiv.org/abs/2606.28551"}, {"kind": "website", "url": "https://www.datacomp.ai/dcvlm/"}], "authors": ["Matteo Farina", "Vishaal Udandarao", "Thao Nguyen", "Selim Kuzucu", "Maximilian Böther", "Andreas Hochlehnert", "Adhiraj Ghosh", "Marianna Nezhurina", "Karsten Roth", "Joschka Struber", "Yuhui Zhang", "Sebastian Dziadzio", "Elaine Sui", "Soumya Jahagirdar", "Dhruba Ghosh", "Hasan Hammoud", "Thomas De Min", "Simone Caldarella", "Jehanzeb Mirza", "Sedrick Keh", "Mehdi Cherti", "Hilde Kuehne", "Bernt Schiele", "Serena Yeung-Levy", "Muhammad Ferjad Naeem", "Federico Tombari", "Ana Klimovic", "Elisa Ricci", "Matthias Bethge", "Sewoong Oh", "Ameya Prabhu", "Alessio Tonioni", "Jenia Jitsev", "Massimiliano Mancini", "Ludwig Schmidt", "Nikhil Parthasarathy"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.28551", "source_url": "https://arxiv.org/abs/2606.28551"}, "categories": ["General AI", "Multimodal"], "description": {"en": "A controlled benchmark for comparing vision-language training-data curation strategies across specified model scales and token budgets."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.28551", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.28551", "title": "DataComp for VLMs (DCVLM)"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2606.28551", "modality": null, "name": "DataComp for VLMs (DCVLM)", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-09-25T04:40:22Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_datacomp-vlm-improved-open-datasets-for-vi_2f0357d4", "origin_source": "arxiv", "recognition_confidence": "0.7", "record_sha256": "6f2d4ab327409bd214c4743fc1833ffe6c112e73f7932e8d29877c559e6d7d18", "relation": "introduces", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T17:21:45.202648Z", "source_url": "https://arxiv.org/abs/2606.28551"}, "publisher": null, "released": "2026-06-26", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.28551", "source_url": "https://arxiv.org/abs/2606.28551"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-28551", "source": "claire_radar", "source_benchmark_id": "2606.28551", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["DataComp-VLM: Improved Open Datasets for Vision-Language Models"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Multimodal", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The large training corpus, broad downstream suite, and public DataComp infrastructure may attract near-term attention from vision-language data researchers.", "score": 72}, "capabilities": [], "confidence": "Medium", "construction": "Transform Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T17:21:45.202648Z", "inputHash": "93792ec73de19b4511fff55a33c91b5a16d5c55ff5f40c8c0f3152abf78e5c43", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "DataComp for VLMs (DCVLM), a benchmark", "canonicalNameSource": "abstract", "decisionReason": "The abstract formally names the benchmark, defines repeatable experimental variables and a scored core suite, and supplies a public artifact path.", "evaluationEvidence": {"reuse": "DCVLM and all accompanying artifacts will be made publicly available at https://www.datacomp.ai/dcvlm/.", "scoring": "63.6% accuracy on our 33-task core suite", "task": "test curation strategies (filtering, mixing, formatting, sampling) across 1B-8B models and 6.25B-200B token budgets"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T17:21:45.202648Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A controlled benchmark for comparing vision-language training-data curation strategies across specified model scales and token budgets.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "We introduce DataComp for VLMs (DCVLM), a benchmark for controlled data-centric experiments to improve VLM training."}, "familyId": "bmf_15e3e323894c", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_datacomp-vlm-improved-open-datasets-for-vi_2f0357d4", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2606.28551", "pdf": "https://arxiv.org/pdf/2606.28551", "project": "https://www.datacomp.ai/dcvlm/", "report": "https://arxiv.org/abs/2606.28551"}, "metrics": [], "motivation": "Building performant Vision-Language Models (VLMs) requires carefully curating large-scale training datasets, yet the community lacks systematic benchmarks for evaluating such curation strategies.", "name": "DataComp for VLMs (DCVLM)", "oneLine": "A controlled benchmark for comparing vision-language training-data curation strategies across specified model scales and token budgets.", "paperTitle": "DataComp-VLM: Improved Open Datasets for Vision-Language Models", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.28551", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Paper only", "recognitionConfidence": 0.7, "relation": "introduces", "releasedAt": "2026-06-26", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Matteo Farina", "Vishaal Udandarao", "Thao Nguyen", "Selim Kuzucu", "Maximilian B\\\"other", "Andreas Hochlehnert", "Adhiraj Ghosh", "Marianna Nezhurina", "Karsten Roth", "Joschka Struber", "Yuhui Zhang", "Sebastian Dziadzio", "Elaine Sui", "Soumya Jahagirdar", "Dhruba Ghosh", "Hasan Hammoud", "Thomas De Min", "Simone Caldarella", "Jehanzeb Mirza", "Sedrick Keh", "Mehdi Cherti", "Hilde Kuehne", "Bernt Schiele", "Serena Yeung-Levy", "Muhammad Ferjad Naeem", "Federico Tombari", "Ana Klimovic", "Elisa Ricci", "Matthias Bethge", "Sewoong Oh", "Ameya Prabhu", "Alessio Tonioni", "Jenia Jitsev", "Massimiliano Mancini", "Ludwig Schmidt", "Nikhil Parthasarathy"], "categories": ["cs.CV", "cs.CL", "cs.LG"], "id": "2606.28551", "title": "DataComp-VLM: Improved Open Datasets for Vision-Language Models", "type": "arxiv", "url": "https://arxiv.org/abs/2606.28551"}, "sourceUpdatedAt": "2026-08-10T01:17:41Z", "topics": ["Multimodal"], "venueAttempts": [], "whyItMatters": "It provides a common experimental framework for isolating how data filtering, mixing, formatting, and sampling affect vision-language model performance."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
