Update croissant.json

croissant.json

@@ -1,38 +1,131 @@
 {
-    "@context": "https://schema.org",
-    "@type": "Dataset",
-    "name": "mipbench",
-    "description": "\n\t\n\t\t\n\t\tMIPBench\n\t\n\nMIPBench is an evaluation-only benchmark for measuring position sensitivity in position-invariant multi-image visual question answering. Each example contains multiple images, a question, an answer, and provenance metadata. The benchmark is intended to evaluate whether a vision-language model changes its answer when the input images are permuted while the ground-truth answer remains unchanged.\n\n\t\n\t\t\n\t\tDataset Summary\n\t\n\nMIPBench contains 6,994 position-invariant multi-image… See the full description on the dataset page: https://huggingface.co/datasets/mipbench-anon/mipbench.",
-    "alternateName": [
-        "mipbench-anon/mipbench",
-        "MIPBench"
-    ],
-    "creator": {
-        "@type": "Person",
-        "name": "mipbench-anon",
-        "url": "https://huggingface.co/mipbench-anon"
-    },
-    "keywords": [
-        "visual-question-answering",
-        "English",
-        "cc-by-nc-4.0",
-        "1K - 10K",
-        "parquet",
-        "optimized-parquet",
-        "Image",
-        "Text",
-        "Datasets",
-        "Dask",
-        "Polars",
-        "Croissant",
-        "🇺🇸 Region: US",
-        "vision-language-models",
-        "multi-image",
-        "benchmark",
-        "evaluation",
-        "position-bias",
-        "visual-question-answering"
-    ],
-    "license": "https://choosealicense.com/licenses/cc-by-nc-4.0/",
-    "url": "https://huggingface.co/datasets/mipbench-anon/mipbench"
+  "@context": {
+    "@language": "en",
+    "@vocab": "https://schema.org/",
+    "sc": "https://schema.org/",
+    "rai": "http://mlcommons.org/croissant/RAI/",
+    "prov": "http://www.w3.org/ns/prov#"
+  },
+  "@type": "Dataset",
+  "name": "mipbench",
+  "description": "\n\t\n\t\t\n\t\tMIPBench\n\t\n\nMIPBench is an evaluation-only benchmark for measuring position sensitivity in position-invariant multi-image visual question answering. Each example contains multiple images, a question, an answer, and provenance metadata. The benchmark is intended to evaluate whether a vision-language model changes its answer when the input images are permuted while the ground-truth answer remains unchanged.\n\n\t\n\t\t\n\t\tDataset Summary\n\t\n\nMIPBench contains 6,994 position-invariant multi-image… See the full description on the dataset page: https://huggingface.co/datasets/mipbench-anon/mipbench.",
+  "alternateName": [
+    "mipbench-anon/mipbench",
+    "MIPBench"
+  ],
+  "creator": {
+    "@type": "Person",
+    "name": "mipbench-anon",
+    "url": "https://huggingface.co/mipbench-anon"
+  },
+  "keywords": [
+    "visual-question-answering",
+    "English",
+    "cc-by-nc-4.0",
+    "1K - 10K",
+    "parquet",
+    "optimized-parquet",
+    "Image",
+    "Text",
+    "Datasets",
+    "Dask",
+    "Polars",
+    "Croissant",
+    "🇺🇸 Region: US",
+    "vision-language-models",
+    "multi-image",
+    "benchmark",
+    "evaluation",
+    "position-bias",
+    "visual-question-answering"
+  ],
+  "license": "https://choosealicense.com/licenses/cc-by-nc-4.0/",
+  "url": "https://huggingface.co/datasets/mipbench-anon/mipbench",
+  "rai:dataLimitations": "English-only benchmark of multi-image VQA samples where the correct answer should not depend on image order. Excludes video, action sequencing, document QA, and text-needle retrieval. For samples with four or more images, only a subset of permutations is evaluated, so reported sensitivity numbers are lower bounds rather than worst-case rates. Not recommended for training, fine-tuning, safety-critical deployment assessment, or cross-lingual evaluation.",
+  "rai:dataBiases": "Image content, language, and demographic coverage are inherited from the source benchmarks and skew toward Western and East Asian scenes with English-only text. Answer formats are imbalanced toward yes/no questions. Long-context samples (10+ images) come almost entirely from one source, so context-length analyses are confounded with that source's distribution. The filter targets surface lexical cues, so position-dependent tasks lacking ordinal wording may be over-represented.",
+  "rai:personalSensitiveInformation": "Inherited from source benchmarks: Geography, Language (English only), Culture, and incidental Gender and Age via natural photographs containing identifiable people. The dataset does not contain health, medical, political, religious, socio-economic, or seniority data. No new personal information was collected.",
+  "rai:dataUseCases": "The dataset measures permutation sensitivity in multi-image vision-language models on tasks whose correct answer is independent of image order. Validated for diagnostic evaluation, stratified analysis across task categories, and target-position accuracy analysis in needle-style retrieval. Not validated for clinical or safety-critical deployment, training data, evaluation of order-dependent tasks, or cross-lingual evaluation.",
+  "rai:dataSocialImpact": "Positive: enables controlled measurement of a multi-image failure mode that prior benchmarks do not isolate, supporting more honest reporting of model behavior under input variation. Risks: misuse as training data, over-interpretation of reported sensitivity as worst-case bounds, and treating low sensitivity as a proxy for deployment reliability. Mitigations: explicit intended-use and out-of-scope statements in the dataset card, per-sample provenance preserved for auditability, and license terms inherited and documented per sample.",
+  "rai:hasSyntheticData": false,
+  "prov:wasDerivedFrom": [
+    {
+      "@id": "https://huggingface.co/datasets/TIGER-Lab/Mantis-Eval",
+      "prov:label": "Mantis-Eval",
+      "sc:license": "Apache-2.0"
+    },
+    {
+      "@id": "https://github.com/pgtrece/DAB",
+      "prov:label": "MIHBench",
+      "sc:license": "Unspecified"
+    },
+    {
+      "@id": "https://huggingface.co/datasets/FreedomIntelligence/MileBench",
+      "prov:label": "MileBench",
+      "sc:license": "CC-BY-2.0"
+    },
+    {
+      "@id": "https://huggingface.co/datasets/VLLMs/MIRB",
+      "prov:label": "MIRB",
+      "sc:license": "MIT"
+    },
+    {
+      "@id": "https://huggingface.co/datasets/FanqingM/MMIU-Benchmark",
+      "prov:label": "MMIU-Benchmark",
+      "sc:license": "CC-BY-4.0"
+    },
+    {
+      "@id": "https://huggingface.co/datasets/RunsenXu/MMSI-Bench",
+      "prov:label": "MMSI-Bench",
+      "sc:license": "CC-BY-4.0"
+    },
+    {
+      "@id": "https://huggingface.co/datasets/MUIRBENCH/MUIRBENCH",
+      "prov:label": "MuirBench",
+      "sc:license": "CC-BY-4.0"
+    },
+    {
+      "@id": "https://huggingface.co/datasets/mehrankazemi/ReMI",
+      "prov:label": "ReMI",
+      "sc:license": "CC"
+    },
+    {
+      "@id": "https://huggingface.co/datasets/tsunghanwu/visual_haystacks",
+      "prov:label": "Visual Haystacks",
+      "sc:license": "CC-BY-4.0"
+    }
+  ],
+  "prov:wasGeneratedBy": [
+    {
+      "@type": "prov:Activity",
+      "prov:type": {
+        "@id": "https://www.wikidata.org/wiki/Q4929239"
+      },
+      "prov:label": "Source Aggregation",
+      "sc:description": "Aggregated multi-image VQA samples from nine source benchmarks. No new images or questions authored; per-sample provenance preserved."
+    },
+    {
+      "@type": "prov:Activity",
+      "prov:type": {
+        "@id": "https://www.wikidata.org/wiki/Q1172378"
+      },
+      "prov:label": "Position-Dependence Filtering",
+      "sc:description": "A three-stage rule-based filter applied a task-level blacklist, regex matching for ordinal, temporal, and per-position markers, and structural detection of role-assigning text between image tokens."
+    },
+    {
+      "@type": "prov:Activity",
+      "prov:type": {
+        "@id": "https://www.wikidata.org/wiki/Q109719325"
+      },
+      "prov:label": "Taxonomy Assignment",
+      "sc:description": "Samples passing the filter were assigned by the authors to one of five taxonomy categories using source-task identifiers and documented boundary rules. No LLM-based annotators were used."
+    },
+    {
+      "@type": "prov:Activity",
+      "prov:type": {
+        "@id": "https://www.wikidata.org/wiki/Q3306762"
+      },
+      "prov:label": "Quality Review and Balancing",
+      "sc:description": "The authors inspected samples passing the filter and removed cases with implicit position dependence that the rules did not catch. Per-category stratified subsampling was then applied toward a common target size, with the over-represented retrieval source capped and large reasoning subtasks reduced proportionally."
+    }
+  ]
 }