Datasets:

anomyousMMDR
/

MMdeepresearch

+{
+  "@context": {
+    "@language": "en",
+    "@vocab": "https://schema.org/",
+    "cr": "http://mlcommons.org/croissant/",
+    "sc": "https://schema.org/",
+    "dct": "http://purl.org/dc/terms/",
+    "citeAs": "cr:citeAs",
+    "column": "cr:column",
+    "conformsTo": "dct:conformsTo",
+    "data": {"@id": "cr:data", "@type": "@json"},
+    "dataCollection": "cr:dataCollection",
+    "dataBiases": "cr:dataBiases",
+    "personalSensitiveInformation": "cr:personalSensitiveInformation",
+    "dataType": {"@id": "cr:dataType", "@type": "@vocab"},
+    "extract": "cr:extract",
+    "field": "cr:field",
+    "fileObject": "cr:fileObject",
+    "fileSet": "cr:fileSet",
+    "format": "cr:format",
+    "includes": "cr:includes",
+    "isArray": "cr:isArray",
+    "arrayShape": "cr:arrayShape",
+    "jsonPath": "cr:jsonPath",
+    "key": "cr:key",
+    "path": "cr:path",
+    "recordSet": "cr:recordSet",
+    "references": "cr:references",
+    "regex": "cr:regex",
+    "source": "cr:source",
+    "transform": "cr:transform",
+    "useCases": "cr:useCases",
+    "disallowedUses": "cr:disallowedUses"
+  },
+  "@type": "sc:Dataset",
+  "conformsTo": "http://mlcommons.org/croissant/1.1",
+  "name": "MMdeepresearch",
+  "alternateName": ["MMDR-Bench", "MMDeepResearch-Bench"],
+  "description": "MMDR-Bench is a benchmark for end-to-end multimodal deep research agents, comprising 140 expert-crafted tasks across 19 domains and two regimes (Daily: 40 tasks across 11 domains, focused on loosely-structured visual inputs such as screenshots and UI captures; Research: 100 tasks across 10 domains, focused on information-dense visuals such as charts, diagrams, and tables). Each task is an image-text bundle that evaluates an agent's ability to read provided visual inputs, cite retrieved web evidence, and produce a citation-grounded multimodal research report. MMDR-Bench is designed for process-oriented evaluation: there is no fixed gold report, because deep-research outputs are inherently non-stationary. The benchmark instead instruments three reference-free process signals (VEF / MOSAIC / TRACE) defined in the accompanying paper.",
+  "license": "https://opensource.org/licenses/MIT",
+  "url": "https://huggingface.co/datasets/anomyousMMDR/MMdeepresearch",
+  "creator": {"@type": "Person", "name": "Anonymous (under double-blind review)"},
+  "citeAs": "Anonymized during peer review. Citation will be provided upon acceptance.",
+  "keywords": [
+    "multimodal", "deep-research", "benchmark", "agent", "evaluation",
+    "citation-grounded", "visual-evidence-fidelity", "process-oriented"
+  ],
+  "dataCollection": "Tasks were proposed and iteratively refined by doctoral-level domain experts. Each candidate task went through three checks before inclusion: (1) clarity, (2) multimodal necessity (the task cannot be solved from text alone), and (3) evidence-grounding (the answer is verifiable via citations). Images paired with each task are drawn from publicly available sources; licensing was audited per image and tasks whose images lacked permissive terms were excluded or replaced with licensed substitutes. No crowdsourcing platform was used; all annotation was by vetted domain experts recruited through academic channels. Compensation followed standard academic research-assistant rates in the relevant jurisdiction and met the minimum-wage guideline of the NeurIPS Code of Ethics.",
+  "dataBiases": "The task-language distribution is strongly skewed toward English and Chinese, with French, German, and Finnish present in the long tail. Domain coverage is 19 fields grouped into Humanities/Cultural Studies, Social/Policy Studies, Economics/Business, Environment/Energy, Life/Health Sciences, Mathematics/Engineering, Computer/Data Science, Interdisciplinary, and an exploratory category; coverage within each field is not exhaustive and reflects the constructing experts' research interests. The task set is static at 140 items and is therefore not representative of all possible multimodal deep-research workflows; rankings should be interpreted as capability snapshots under this specific probe. Because images were selected by experts, there is an implicit aesthetic/epistemic prior toward figures, charts, and documents that are legible to trained researchers. Users should not treat absolute scores as cross-time comparable because both the evaluated models and the judge LLM are accessed via external APIs whose snapshots drift.",
+  "personalSensitiveInformation": "Tasks and images were screened to exclude personally identifiable information (PII). No medical records, government identifiers, biometric data, or private communications are included. Publicly visible faces that appear in historical or institutional photographs were retained only when the underlying source license permitted redistribution; no attempt is made to link faces to identities. Users of the dataset must not attempt re-identification of individuals who may incidentally appear in images. A takedown channel is available for any subject who later requests removal; in such cases the affected task is redacted and the dataset version is bumped.",
+  "useCases": "Intended uses: (i) benchmarking multimodal deep-research agents and the LLMs that power them; (ii) measuring whether an agent grounds its textual claims in retrieved web evidence (TRACE) and in the provided visual inputs (VEF); (iii) measuring internal consistency between a generated report's visual-referenced statements and its own visual artifacts (MOSAIC); (iv) studying the evaluation methodology itself, including judge consistency and human-alignment studies.",
+  "disallowedUses": "Not intended for: (i) training or fine-tuning models on the task pool in ways that would contaminate future evaluations; (ii) deployment-readiness certification of any commercial deep-research product (the benchmark certifies groundedness, not absolute factual correctness); (iii) any downstream task whose correctness depends on absolute scores being comparable across time, because both model and judge APIs drift.",
+  "distribution": [
+    {
+      "@type": "cr:FileObject",
+      "@id": "repo",
+      "name": "repo",
+      "description": "Anonymous Hugging Face dataset repository (review-period mirror).",
+      "contentUrl": "https://huggingface.co/datasets/anomyousMMDR/MMdeepresearch/tree/refs%2Fconvert%2Fparquet",
+      "encodingFormat": "git+https",
+      "sha256": "https://github.com/mlcommons/croissant/issues/80"
+    },
+    {
+      "@type": "cr:FileSet",
+      "@id": "parquet-files-for-config-default",
+      "containedIn": {"@id": "repo"},
+      "encodingFormat": "application/x-parquet",
+      "includes": "default/*/*.parquet"
+    }
+  ],
+  "recordSet": [
+    {
+      "@type": "cr:RecordSet",
+      "@id": "default_splits",
+      "name": "default_splits",
+      "description": "Splits for the default config.",
+      "dataType": "cr:Split",
+      "key": {"@id": "default_splits/split_name"},
+      "field": [{"@type": "cr:Field", "@id": "default_splits/split_name", "dataType": "sc:Text"}],
+      "data": [{"default_splits/split_name": "train"}]
+    },
+    {
+      "@type": "cr:RecordSet",
+      "@id": "default",
+      "description": "The 140-task MMDR-Bench records.",
+      "field": [
+        {"@type": "cr:Field", "@id": "default/split", "dataType": "sc:Text",
+          "source": {"fileSet": {"@id": "parquet-files-for-config-default"},
+            "extract": {"fileProperty": "fullpath"},
+            "transform": {"regex": "default/(?:partial-)?(train)/.+parquet$"}},
+          "references": {"field": {"@id": "default_splits/split_name"}}},
+        {"@type": "cr:Field", "@id": "default/id", "dataType": "cr:Int64",
+          "source": {"fileSet": {"@id": "parquet-files-for-config-default"}, "extract": {"column": "id"}}},
+        {"@type": "cr:Field", "@id": "default/caption", "dataType": "sc:Text",
+          "source": {"fileSet": {"@id": "parquet-files-for-config-default"}, "extract": {"column": "caption"}}},
+        {"@type": "cr:Field", "@id": "default/body", "dataType": "sc:Text",
+          "source": {"fileSet": {"@id": "parquet-files-for-config-default"}, "extract": {"column": "body"}}},
+        {"@type": "cr:Field", "@id": "default/image_url", "dataType": "sc:ImageObject",
+          "source": {"fileSet": {"@id": "parquet-files-for-config-default"}, "extract": {"column": "image_url"},
+            "transform": {"jsonPath": "bytes"}},
+          "isArray": true, "arrayShape": "-1"},
+        {"@type": "cr:Field", "@id": "default/tags", "dataType": "sc:Text",
+          "source": {"fileSet": {"@id": "parquet-files-for-config-default"}, "extract": {"column": "tags"}},
+          "isArray": true, "arrayShape": "-1"},
+        {"@type": "cr:Field", "@id": "default/language", "dataType": "sc:Text",
+          "source": {"fileSet": {"@id": "parquet-files-for-config-default"}, "extract": {"column": "language"}}},
+        {"@type": "cr:Field", "@id": "default/difficulty", "dataType": "sc:Text",
+          "source": {"fileSet": {"@id": "parquet-files-for-config-default"}, "extract": {"column": "difficulty"}}}
+      ]
+    }
+  ]
+}