{ "@context": { "@language": "en", "@vocab": "https://schema.org/", "citeAs": "cr:citeAs", "column": "cr:column", "conformsTo": "dct:conformsTo", "cr": "http://mlcommons.org/croissant/", "rai": "http://mlcommons.org/croissant/RAI/", "data": {"@id": "cr:data", "@type": "@json"}, "dataType": {"@id": "cr:dataType", "@type": "@vocab"}, "dct": "http://purl.org/dc/terms/", "equivalentProperty": "cr:equivalentProperty", "examples": {"@id": "cr:examples", "@type": "@json"}, "extract": "cr:extract", "field": "cr:field", "fileObject": "cr:fileObject", "fileProperty": "cr:fileProperty", "fileSet": "cr:fileSet", "format": "cr:format", "includes": "cr:includes", "isLiveDataset": "cr:isLiveDataset", "jsonPath": "cr:jsonPath", "key": "cr:key", "md5": "cr:md5", "parentField": "cr:parentField", "path": "cr:path", "recordSet": "cr:recordSet", "references": "cr:references", "regex": "cr:regex", "repeated": "cr:repeated", "replace": "cr:replace", "samplingRate": "cr:samplingRate", "sc": "https://schema.org/", "separator": "cr:separator", "source": "cr:source", "subField": "cr:subField", "transform": "cr:transform" }, "@type": "sc:Dataset", "name": "MemeBench", "description": "MemeBench is a bilingual diagnostic benchmark for open-ended meme interpretation, evaluating large vision-language models' cultural-semantic understanding. It contains 1,253 memes (768 Chinese, 485 English) annotated under the VIKR schema, which decomposes meme understanding into four layers: Visual clues (V), Identity links (I), Knowledge units (K), and Reasoning mechanisms (R). Each meme is paired with per-dimension evaluation checklists for layer-wise diagnostic analysis.", "conformsTo": "http://mlcommons.org/croissant/1.0", "url": "https://huggingface.co/datasets/anonymous-neurips-2026/memebench", "version": "1.0.0", "datePublished": "2026-05-01", "license": "https://creativecommons.org/licenses/by-nc-sa/4.0/", "creator": { "@type": "sc:Organization", "name": "Anonymous Authors (NeurIPS 2026 ED Track Submission)" }, "keywords": [ "diagnostic benchmark", "vision-language models", "cultural understanding", "multimodal evaluation", "meme interpretation", "VIKR schema" ], "inLanguage": ["zh", "en"], "rai:dataCollection": "Memes were collected via web scraping from Chinese and English internet communities (forums, social media, image boards). Annotation followed a multi-stage pipeline: (1) LLM-assisted initial annotation generating structured VIKR fields, (2) human expert verification and correction of all entries, (3) automated quality audits enforcing schema invariants (entity dual-linkage, cardinality bounds, layer separation). Inter-annotator agreement was measured on a subset. Content moderation flags (political, hate_speech) were applied via automated filters with human override.", "rai:dataLimitations": [ "Domain imbalance: ACG (Anime/Comics/Games) accounts for 50.1% of entries, which may bias model evaluation toward pop-culture knowledge.", "Language imbalance: Chinese entries (61.3%) outnumber English entries (38.7%), reflecting the collection sources.", "This is an evaluation-only benchmark, not intended for training. Models should not be fine-tuned on this data.", "Binary moderation flags (political, hate_speech) provide coarse content safety signals but do not capture nuanced sensitivity." ], "rai:dataBiases": [ "Domain bias: ACG-dominant (50.1%), with Sports (0.6%) and Politics (0.5%) severely underrepresented.", "Cultural bias: Memes are drawn primarily from Chinese internet culture (Bilibili, Weibo) and English-language platforms, excluding other cultural contexts.", "Temporal bias: Memes were collected in 2025-2026 and reflect contemporary internet culture; older or emerging memes may not be represented.", "Logic type bias: Identity Collision (43.8%) and Irony (29.0%) dominate; Visual Similarity (2.5%) and Phonetic Pun (3.7%) are rare." ], "rai:dataSocialImpact": "MemeBench enables research on AI understanding of cultural communication through memes. Memes are a significant form of online discourse, and improving AI comprehension of memes can benefit content moderation, accessibility, and cross-cultural communication. Some memes reference sensitive cultural or political topics in satirical contexts. The benchmark does not contain hate speech or politically extreme content (all such items were filtered during curation).", "rai:personalSensitiveInformation": "The dataset contains no personally identifiable information (PII). All memes are sourced from publicly available internet content. Some entries reference public figures (celebrities, fictional characters, historical figures) in satirical or humorous contexts, consistent with fair use in cultural commentary. No private individuals are depicted or identified.", "rai:dataAnnotationProtocol": "Each meme is annotated under the VIKR schema through a multi-stage pipeline: (1) a primary annotator writes a free-text ground-truth explanation, which is converted into structured VIKR fields (visual description, identity entities, knowledge facts, reasoning mechanism) via an LLM-assisted pipeline using Gemini-3-Pro; (2) all entries are reviewed and corrected by domain experts; (3) automated quality audits enforce schema invariants including entity dual-linkage between visual and identity layers, cardinality bounds (1-4 items per checklist/facts), and layer separation constraints.", "rai:annotationsPerItem": "1 primary annotator + 1 auditor per item. Inter-annotator agreement measured on a 150-item stratified subset: logic type (Krippendorff alpha=0.84), domain (alpha=0.83), visual structure (alpha=0.81), meme type (alpha=0.80).", "rai:dataPreprocessingProtocol": "Two-stage filtering from 1,500 candidate memes to 1,253 retained items. Stage 1 (moderation): automated content filter removes political and hate-speech content (244 items removed, 95% of Politics-domain memes). Stage 2 (coverage): items with no model coverage across all tested LVLMs are excluded (34 items). The filtering disproportionately removes English political memes but does not introduce further domain bias in Stage 2.", "rai:dataUseCases": [ "Diagnostic evaluation of large vision-language models on cultural-semantic meme understanding.", "Layer-wise analysis of model failures across Visual, Identity, Knowledge, and Reasoning dimensions.", "Benchmarking retrieval-augmented methods for culturally grounded multimodal reasoning.", "Cross-lingual comparison of model performance on Chinese and English internet memes." ], "rai:hasSyntheticData": "The meme images are not synthetic; all are collected from publicly available internet sources. The structured VIKR annotations were initially generated via an LLM-assisted pipeline (Gemini-3-Pro) from human-written reference explanations, then reviewed and corrected by human experts. Evaluation checklists are derived from these human-verified annotations.", "citeAs": "@inproceedings{anonymous2026memebench, title={MemeBench: Diagnosing Cultural-Semantic Understanding in LVLMs through Memes}, author={Anonymous}, booktitle={NeurIPS 2026 Evaluations and Datasets Track}, year={2026}}", "distribution": [ { "@type": "cr:FileObject", "@id": "repo", "name": "repo", "description": "The HuggingFace git repository.", "contentUrl": "https://huggingface.co/datasets/anonymous-neurips-2026/memebench/tree/main", "encodingFormat": "git+https", "sha256": "https://github.com/mlcommons/croissant/issues/80" }, { "@type": "cr:FileObject", "@id": "annotations-file", "name": "memebench_v1.json", "description": "JSON file containing 1,253 annotated meme entries with VIKR schema.", "contentUrl": "https://huggingface.co/datasets/anonymous-neurips-2026/memebench/resolve/main/memebench_v1.json", "encodingFormat": "application/json", "sha256": "162a7d1008d45acebc5ccf00627b0240bf625964f9fa44a11aad7ae545923d21" }, { "@type": "cr:FileSet", "@id": "image-files", "name": "images", "description": "PNG images of 1,253 internet memes.", "encodingFormat": "image/png", "containedIn": {"@id": "repo"}, "includes": "images/*.png" } ], "recordSet": [ { "@type": "cr:RecordSet", "@id": "memes", "name": "memes", "description": "Each record represents one annotated meme with VIKR evaluation dimensions.", "field": [ { "@type": "cr:Field", "@id": "memes/id", "name": "id", "description": "Unique integer identifier for the meme.", "dataType": "sc:Integer", "source": { "fileObject": {"@id": "annotations-file"}, "extract": {"jsonPath": "$[*].id"} } }, { "@type": "cr:Field", "@id": "memes/image_path", "name": "image_path", "description": "Relative path to the meme image file.", "dataType": "sc:Text", "source": { "fileObject": {"@id": "annotations-file"}, "extract": {"jsonPath": "$[*].image_path"} } }, { "@type": "cr:Field", "@id": "memes/checked_gt", "name": "checked_gt", "description": "Human-verified ground truth explanation of the meme.", "dataType": "sc:Text", "source": { "fileObject": {"@id": "annotations-file"}, "extract": {"jsonPath": "$[*].checked_gt"} } }, { "@type": "cr:Field", "@id": "memes/domain", "name": "domain", "description": "Cultural domain: ACG, Movies_TV, History, Politics, Daily_Life, Sports, or Cross_Domain.", "dataType": "sc:Text", "source": { "fileObject": {"@id": "annotations-file"}, "extract": {"jsonPath": "$[*].meta.domain"} } }, { "@type": "cr:Field", "@id": "memes/language", "name": "language", "description": "Primary language of the meme: zh or en.", "dataType": "sc:Text", "source": { "fileObject": {"@id": "annotations-file"}, "extract": {"jsonPath": "$[*].meta.language"} } }, { "@type": "cr:Field", "@id": "memes/meme_type", "name": "meme_type", "description": "Meme type: Single_Source, Multi_Source, or Template.", "dataType": "sc:Text", "source": { "fileObject": {"@id": "annotations-file"}, "extract": {"jsonPath": "$[*].meta.meme_type"} } }, { "@type": "cr:Field", "@id": "memes/structure", "name": "structure", "description": "Visual structure: Single, Comparison, Sequential, or Collage.", "dataType": "sc:Text", "source": { "fileObject": {"@id": "annotations-file"}, "extract": {"jsonPath": "$[*].meta.structure"} } }, { "@type": "cr:Field", "@id": "memes/logic_type", "name": "logic_type", "description": "Primary humor mechanism: VS (Visual Similarity), PP (Phonetic Pun), SM (Semantic Mismatch), IC (Identity Collision), RV (Rule Violation), or IR (Irony).", "dataType": "sc:Text", "source": { "fileObject": {"@id": "annotations-file"}, "extract": {"jsonPath": "$[*].reasoning.logic"} } } ] } ] }