Add Croissant 1.0 metadata with RAI extension (mlcroissant-validated); corrected QC stats (2.3% flagged, 31.3% adjudicated)

cei_benchmark.croissant.json

@@ -0,0 +1,233 @@
+{
+  "@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",
+    "samplingRate": "cr:samplingRate",
+    "examples": {
+      "@id": "cr:examples",
+      "@type": "@json"
+    },
+    "extract": "cr:extract",
+    "field": "cr:field",
+    "fileProperty": "cr:fileProperty",
+    "fileObject": "cr:fileObject",
+    "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",
+    "sc": "https://schema.org/",
+    "separator": "cr:separator",
+    "source": "cr:source",
+    "subField": "cr:subField",
+    "transform": "cr:transform"
+  },
+  "@type": "sc:Dataset",
+  "conformsTo": "http://mlcommons.org/croissant/1.0",
+  "name": "CEI-Benchmark",
+  "alternateName": "Contextual Emotional Inference Benchmark",
+  "description": "300 expert-authored scenarios for evaluating pragmatic emotion inference in language models. The benchmark covers five pragmatic subtypes (sarcasm/irony, mixed signals, strategic politeness, passive aggression, deflection/misdirection) crossed with three speaker-listener power relations (peer, superior-to-subordinate, subordinate-to-superior). Each scenario received three independent annotations from 15 trained annotators using Plutchik's 8-emotion taxonomy and Valence-Arousal-Dominance ratings (900 annotations total). Per-annotator labels are released alongside adjudicated gold labels so that human disagreement can be modeled rather than collapsed.",
+  "license": "https://creativecommons.org/licenses/by/4.0/",
+  "url": "https://github.com/jon-chun/cei-tom-dataset-base",
+  "sameAs": [
+    "https://huggingface.co/datasets/jonc/cei-benchmark",
+    "https://doi.org/10.5281/zenodo.18528705"
+  ],
+  "version": "1.0.0",
+  "datePublished": "2026-02-01",
+  "citeAs": "Chun et al. (2026). CEI: A Benchmark for Evaluating Pragmatic Reasoning in Language Models. DMLR submission #177; arXiv:2603.09993.",
+  "keywords": [
+    "pragmatic reasoning",
+    "benchmark",
+    "language models",
+    "emotion inference",
+    "social reasoning",
+    "power dynamics",
+    "theory of mind",
+    "annotation methodology"
+  ],
+  "creator": {
+    "@type": "sc:Person",
+    "name": "Jon Chun",
+    "affiliation": "Kenyon College",
+    "email": "chunj@kenyon.edu"
+  },
+  "contributor": {
+    "@type": "sc:Organization",
+    "name": "CEI author team (Kenyon College and Yale University)"
+  },
+  "rai:dataCollection": "Expert authoring by two lead researchers following a structured template (context, roles, utterance, subtype tag); each scenario independently reviewed by a second researcher for clarity, subtype fidelity, and cultural accessibility. 600 initial candidates were narrowed to 300 after manual quality review for balance across subtypes and power relations. No web scraping; no real conversations; all scenarios synthetic.",
+  "rai:dataCollectionType": "Manual curation (expert-authored synthetic scenarios)",
+  "rai:dataBiases": "Scenarios skew toward Western workplace and household contexts. Power-relation distribution is 72% peer / 20% high-to-low / 7% low-to-high, mirroring everyday communication but limiting power-stratified analyses. Scenario authorship by two researchers and annotation by 15 undergraduate students at a single US institution (Kenyon College) may limit generalizability; agreement values may not transfer to populations with different cultural or professional backgrounds. English-only; not calibrated for non-Western communicative norms.",
+  "rai:dataLimitations": "NOT for training emotion-recognition models (deliberate ambiguity is a feature, not noise); NOT for clinical alexithymia diagnosis or any clinical decision support; NOT calibrated for non-Western communicative norms or non-English languages; NOT for production HR screening, workplace surveillance, or automated employee sentiment analysis; NOT a substitute for context-specific validation in any deployed system. Inter-annotator agreement is intentionally low (Fleiss' kappa = 0.21 pooled): pragmatic utterances admit multiple valid readings, and each subtype was rated by a disjoint annotator trio, so per-subtype differences confound phenomenon difficulty with trio effects.",
+  "rai:dataUseCases": "Evaluation of pragmatic reasoning in large language models; RSA-baseline benchmarking; annotation-difficulty modeling and disagreement-as-signal research; soft-label and calibration-aware training research; diagnostic analysis of subtype-specific LLM blind spots.",
+  "rai:personalSensitiveInformation": "None. Synthetic workplace, family, social, and service-encounter scenarios; no real-person PII; no real conversations were collected or transcribed. Annotator free-text explanations are withheld from release to protect annotator privacy.",
+  "rai:annotationsPerItem": "3",
+  "rai:totalAnnotators": 15,
+  "rai:annotationPlatform": "Label Studio",
+  "rai:annotatorDemographics": "15 undergraduate students at Kenyon College (interdisciplinary course IPHS 391), all fluent English speakers; trained via two in-class sessions plus written guidelines; IRB-exempt (Category 2); course credit with co-authorship offered for high-quality contributions.",
+  "rai:annotationProcess": "Each subtype received exactly 3 annotators who rated all 60 scenarios in that subtype (900 annotations total). Annotators provided a Plutchik 8-emotion category, Valence-Arousal-Dominance ratings on 7-point scales, a confidence rating, and 2-4 sentence free-text reasoning. A 4-level QC pipeline (schema validation, statistical outlier detection, agreement analysis, expert adjudication) was applied: 2.3% (21/900) of annotations were flagged at the statistical-consistency stage, no annotator was excluded, and 31.3% (94/300) of scenarios (all three-way splits) received expert adjudication.",
+  "rai:dataReleaseMaintenancePlan": "Semantic versioning (v1.0.0). CC-BY-4.0 (data), MIT (code). Hosted on GitHub (primary) and HuggingFace (dataset mirror); persistently archived on Zenodo (concept DOI 10.5281/zenodo.18528705). Maintained for at least 5 years post-publication; deprecation notices and errata via the repository issue tracker and CHANGELOG. Contact: chunj@kenyon.edu.",
+  "distribution": [
+    {
+      "@type": "cr:FileObject",
+      "@id": "data_sarcasm-irony.csv",
+      "name": "data_sarcasm-irony.csv",
+      "description": "60 sarcasm/irony scenarios with context, roles, utterance, three per-annotator Plutchik + VAD + confidence labels, gold label, and adjudication notes.",
+      "contentUrl": "data/human-gold/data_sarcasm-irony.csv",
+      "encodingFormat": "text/csv",
+      "sha256": "366041db55f12bcd004a530e766fd12eea6840e4d6b1a34da0b9ed14e556b9f6"
+    },
+    {
+      "@type": "cr:FileObject",
+      "@id": "data_mixed-signals.csv",
+      "name": "data_mixed-signals.csv",
+      "description": "60 mixed-signals scenarios with context, roles, utterance, three per-annotator Plutchik + VAD + confidence labels, gold label, and adjudication notes.",
+      "contentUrl": "data/human-gold/data_mixed-signals.csv",
+      "encodingFormat": "text/csv",
+      "sha256": "88458510d2824be6d968960fc321343d5e4c2ba30c2549db7b9c8876ee6a0911"
+    },
+    {
+      "@type": "cr:FileObject",
+      "@id": "data_strategic-politeness.csv",
+      "name": "data_strategic-politeness.csv",
+      "description": "60 strategic-politeness scenarios with context, roles, utterance, three per-annotator Plutchik + VAD + confidence labels, gold label, and adjudication notes.",
+      "contentUrl": "data/human-gold/data_strategic-politeness.csv",
+      "encodingFormat": "text/csv",
+      "sha256": "76c270f9556e1b933394a5db48ec53f8ca87d538f24718cf12e789b86e13974b"
+    },
+    {
+      "@type": "cr:FileObject",
+      "@id": "data_passive-aggression.csv",
+      "name": "data_passive-aggression.csv",
+      "description": "60 passive-aggression scenarios with context, roles, utterance, three per-annotator Plutchik + VAD + confidence labels, gold label, and adjudication notes.",
+      "contentUrl": "data/human-gold/data_passive-aggression.csv",
+      "encodingFormat": "text/csv",
+      "sha256": "ecf19d3f50e6cd867129f7bb096f9fb7edb899bec5164ea3cb176dba9d4b0028"
+    },
+    {
+      "@type": "cr:FileObject",
+      "@id": "data_deflection-misdirection.csv",
+      "name": "data_deflection-misdirection.csv",
+      "description": "60 deflection/misdirection scenarios with context, roles, utterance, three per-annotator Plutchik + VAD + confidence labels, gold label, and adjudication notes.",
+      "contentUrl": "data/human-gold/data_deflection-misdirection.csv",
+      "encodingFormat": "text/csv",
+      "sha256": "fd81867b7d73ccd515a57e823c0b009efb76f174278854fe7f06ede50fcc6eee"
+    },
+    {
+      "@type": "cr:FileObject",
+      "@id": "splits.json",
+      "name": "splits.json",
+      "description": "Deterministic stratified train/val/test split (70/15/15) balanced by subtype and power relation, seed=42, including the recorded 60-scenario evaluation subset.",
+      "contentUrl": "reports/dmlr2026/splits.json",
+      "encodingFormat": "application/json",
+      "sha256": "a639b35fa413d3183bae16edfd4523d70972dcddec3bf440d8a3ff309fb88af2"
+    }
+  ],
+  "recordSet": [
+    {
+      "@type": "cr:RecordSet",
+      "@id": "sarcasm_irony_records",
+      "name": "sarcasm_irony_records",
+      "description": "Per-scenario records for the sarcasm/irony subtype: id, situational context, speaker/listener roles, utterance, and adjudicated gold label. Per-annotator emotion, VAD, and confidence columns are present in the CSV under per-annotator column names.",
+      "field": [
+        {
+          "@type": "cr:Field",
+          "@id": "sarcasm_irony_records/id",
+          "name": "id",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": { "@id": "data_sarcasm-irony.csv" },
+            "extract": { "column": "id" }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "sarcasm_irony_records/situation",
+          "name": "situation",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": { "@id": "data_sarcasm-irony.csv" },
+            "extract": { "column": "sd_situation" }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "sarcasm_irony_records/speaker_role",
+          "name": "speaker_role",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": { "@id": "data_sarcasm-irony.csv" },
+            "extract": { "column": "sd_speaker_role" }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "sarcasm_irony_records/listener_role",
+          "name": "listener_role",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": { "@id": "data_sarcasm-irony.csv" },
+            "extract": { "column": "sd_listener_role" }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "sarcasm_irony_records/utterance",
+          "name": "utterance",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": { "@id": "data_sarcasm-irony.csv" },
+            "extract": { "column": "sd_utterance" }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "sarcasm_irony_records/gold_standard",
+          "name": "gold_standard",
+          "dataType": "sc:Text",
+          "description": "Adjudicated gold Plutchik emotion: joy, trust, fear, surprise, sadness, disgust, anger, anticipation.",
+          "source": {
+            "fileObject": { "@id": "data_sarcasm-irony.csv" },
+            "extract": { "column": "gold_standard" }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "sarcasm_irony_records/notes",
+          "name": "notes",
+          "dataType": "sc:Text",
+          "description": "Meta-annotator adjudication note where applicable (31.3% of scenarios dataset-wide).",
+          "source": {
+            "fileObject": { "@id": "data_sarcasm-irony.csv" },
+            "extract": { "column": "Notes" }
+          }
+        }
+      ]
+    }
+  ]
+}