TRACE / croissant.jsonld
d4x1's picture taoist's picture
Duplicate from taoist/TRACE
fefcb3a
Raw History Blame Contribute Delete
26.1 kB
{
"@context": {
"@language": "en",
"@vocab": "https://schema.org/",
"arrayShape": "cr:arrayShape",
"citeAs": "cr:citeAs",
"column": "cr:column",
"conformsTo": "dct:conformsTo",
"containedIn": "cr:containedIn",
"cr": "http://mlcommons.org/croissant/",
"rai": "http://mlcommons.org/croissant/RAI/",
"prov": "http://www.w3.org/ns/prov#",
"rdfs": "http://www.w3.org/2000/01/rdf-schema#",
"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"
},
"samplingRate": "cr:samplingRate",
"extract": "cr:extract",
"field": "cr:field",
"fileProperty": "cr:fileProperty",
"fileObject": "cr:fileObject",
"fileSet": "cr:fileSet",
"format": "cr:format",
"includes": "cr:includes",
"isArray": "cr:isArray",
"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",
"distribution": [
{
"@type": "cr:FileObject",
"@id": "repo",
"name": "repo",
"description": "The Hugging Face git repository.",
"contentUrl": "https://huggingface.co/datasets/taoist/TRACE/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",
"dataType": "cr:Split",
"key": {
"@id": "default_splits/split_name"
},
"@id": "default_splits",
"name": "default_splits",
"description": "Splits for the default config.",
"field": [
{
"@type": "cr:Field",
"@id": "default_splits/split_name",
"dataType": "sc:Text"
}
],
"data": [
{
"default_splits/split_name": "train"
},
{
"default_splits/split_name": "validation"
},
{
"default_splits/split_name": "test"
}
]
},
{
"@type": "cr:RecordSet",
"@id": "default",
"description": "taoist/TRACE - 'default' subset\n\nAdditional information:\n- 3 splits: train, validation, test\n- 1 skipped column: turns",
"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|validation|test)/.+parquet$"
}
},
"references": {
"field": {
"@id": "default_splits/split_name"
}
}
},
{
"@type": "cr:Field",
"@id": "default/dialogue_id",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "dialogue_id"
}
}
},
{
"@type": "cr:Field",
"@id": "default/city",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "city"
}
}
},
{
"@type": "cr:Field",
"@id": "default/scenario",
"subField": [
{
"@type": "cr:Field",
"@id": "default/scenario/persona_type",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "scenario"
},
"transform": {
"jsonPath": "persona_type"
}
}
},
{
"@type": "cr:Field",
"@id": "default/scenario/persona_description",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "scenario"
},
"transform": {
"jsonPath": "persona_description"
}
}
},
{
"@type": "cr:Field",
"@id": "default/scenario/primary_poi_type",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "scenario"
},
"transform": {
"jsonPath": "primary_poi_type"
}
}
},
{
"@type": "cr:Field",
"@id": "default/scenario/travel_context",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "scenario"
},
"transform": {
"jsonPath": "travel_context"
}
}
}
]
},
{
"@type": "cr:Field",
"@id": "default/referenced_pois",
"subField": [
{
"@type": "cr:Field",
"@id": "default/referenced_pois/business_id",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "referenced_pois"
},
"transform": {
"jsonPath": "business_id"
}
}
},
{
"@type": "cr:Field",
"@id": "default/referenced_pois/name",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "referenced_pois"
},
"transform": {
"jsonPath": "name"
}
}
},
{
"@type": "cr:Field",
"@id": "default/referenced_pois/poi_type",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "referenced_pois"
},
"transform": {
"jsonPath": "poi_type"
}
}
},
{
"@type": "cr:Field",
"@id": "default/referenced_pois/stars",
"dataType": "cr:Float64",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "referenced_pois"
},
"transform": {
"jsonPath": "stars"
}
}
},
{
"@type": "cr:Field",
"@id": "default/referenced_pois/categories",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "referenced_pois"
},
"transform": {
"jsonPath": "categories"
}
},
"isArray": true,
"arrayShape": "-1"
},
{
"@type": "cr:Field",
"@id": "default/referenced_pois/address",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "referenced_pois"
},
"transform": {
"jsonPath": "address"
}
}
},
{
"@type": "cr:Field",
"@id": "default/referenced_pois/price_range",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "referenced_pois"
},
"transform": {
"jsonPath": "price_range"
}
}
}
],
"isArray": true,
"arrayShape": "-1"
},
{
"@type": "cr:Field",
"@id": "default/metadata",
"subField": [
{
"@type": "cr:Field",
"@id": "default/metadata/model",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "metadata"
},
"transform": {
"jsonPath": "model"
}
}
},
{
"@type": "cr:Field",
"@id": "default/metadata/generated_at",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "metadata"
},
"transform": {
"jsonPath": "generated_at"
}
}
},
{
"@type": "cr:Field",
"@id": "default/metadata/n_candidate_pois",
"dataType": "cr:Int64",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "metadata"
},
"transform": {
"jsonPath": "n_candidate_pois"
}
}
},
{
"@type": "cr:Field",
"@id": "default/metadata/candidate_poi_ids",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "metadata"
},
"transform": {
"jsonPath": "candidate_poi_ids"
}
},
"isArray": true,
"arrayShape": "-1"
},
{
"@type": "cr:Field",
"@id": "default/metadata/open_set_pool_ids",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "metadata"
},
"transform": {
"jsonPath": "open_set_pool_ids"
}
},
"isArray": true,
"arrayShape": "-1"
},
{
"@type": "cr:Field",
"@id": "default/metadata/difficulty",
"dataType": "sc:Text",
"source": {
"fileSet": {
"@id": "parquet-files-for-config-default"
},
"extract": {
"column": "metadata"
},
"transform": {
"jsonPath": "difficulty"
}
}
}
]
}
]
}
],
"conformsTo": "http://mlcommons.org/croissant/1.0",
"name": "TRACE",
"description": "TRACE is a 10,000-dialogue benchmark for conversational tourism recommendation that jointly tests three competencies: (1) recommendation accuracy, (2) review-grounded citation evidence, and (3) rejection recovery. Each multi-turn dialogue between a persona-driven user and an assistant recommender carries verbatim review-span citation pointers (review_id + character offsets) into a shared knowledge base of 2,400 Yelp points of interest (restaurants, attractions, hotels) across 8 U.S. cities and 34,208 reviews. 29.4% of dialogues contain explicit reject_and_refine turns that probe recovery. The release contains 8,001 train / 999 validation / 1,000 test dialogues (stratified by city x POI-type, seed 42), 47 persona archetypes, four difficulty tiers, and an open-set candidate pool (~50-100 POIs per dialogue) alongside the closed 8-POI protocol. Companion artifacts: 14 baseline implementations and a 25-metric evaluation toolkit.",
"alternateName": [
"Tourism Recommendation with Accountable Citation Evidence",
"TRACE Benchmark"
],
"creator": [
{
"@type": "Person",
"name": "Zixu Zhao",
"affiliation": {
"@type": "Organization",
"name": "UNSW Sydney"
}
},
{
"@type": "Person",
"name": "Sijin Wang",
"affiliation": {
"@type": "Organization",
"name": "University of Adelaide"
}
},
{
"@type": "Person",
"name": "Yu Hou",
"affiliation": {
"@type": "Organization",
"name": "Yonsei University"
}
},
{
"@type": "Person",
"name": "Yuanyuan Xu",
"affiliation": {
"@type": "Organization",
"name": "UNSW Sydney"
}
},
{
"@type": "Person",
"name": "Yufan Sheng",
"affiliation": {
"@type": "Organization",
"name": "UNSW Sydney"
}
},
{
"@type": "Person",
"name": "Xike Xie",
"affiliation": {
"@type": "Organization",
"name": "University of Science and Technology of China"
}
},
{
"@type": "Person",
"name": "Wenjie Zhang",
"affiliation": {
"@type": "Organization",
"name": "UNSW Sydney"
}
},
{
"@type": "Person",
"name": "Won-Yong Shin",
"affiliation": {
"@type": "Organization",
"name": "Yonsei University"
}
},
{
"@type": "Person",
"name": "Xin Cao",
"affiliation": {
"@type": "Organization",
"name": "UNSW Sydney"
}
}
],
"keywords": [
"conversational recommendation",
"tourism recommendation",
"point of interest",
"citation grounding",
"review-grounded recommendation",
"rejection recovery",
"multi-turn dialogue",
"open-set evaluation",
"Yelp",
"benchmark",
"datasets",
"natural language processing"
],
"license": "https://creativecommons.org/licenses/by/4.0/",
"url": "https://huggingface.co/datasets/taoist/TRACE",
"sameAs": "https://huggingface.co/datasets/taoist/TRACE",
"version": "1.0.0",
"datePublished": "2026-05-02",
"inLanguage": "en",
"isAccessibleForFree": true,
"usageInfo": "Synthetic content authored by us (dialogue structure, non-quote turn text, grounding annotations, quote_spans positional metadata, code) is released under CC BY 4.0. Verbatim Yelp review text is NOT redistributed: in released dialogue text it appears as [Q:N] placeholders backed by positional metadata (review ID + character offsets), and is hydrated locally from a user-obtained Yelp Open Dataset bundle under the Yelp Dataset User Agreement (non-commercial, non-sublicensable, revocable; https://www.yelp.com/dataset). Yelp-derived POI metadata (name, address, lat/lon, categories, stars, price) is included under the Yelp ToU 4.E academic-publication carve-out. business_id and review_id values appear as bare factual identifiers. See NOTICE.md for the full clause-by-clause split.",
"citeAs": "@inproceedings{zhao2026trace,\n title = {TRACE: Tourism Recommendation with Accountable Citation Evidence},\n author = {Zhao, Zixu},\n year = {2026},\n url = {https://huggingface.co/datasets/taoist/TRACE}\n}",
"rai:hasSyntheticData": true,
"prov:wasGeneratedBy": {
"@type": "prov:Activity",
"rdfs:label": "TRACE synthetic-dialogue generation pipeline",
"description": "Five-stage pipeline orchestrated by DSPy and executed by gpt-5.4-mini: (1) scenario construction conditioned on a persona archetype, (2) candidate POI selection (8 POIs sampled by stratified rating/category from the Yelp Open Dataset), (3) multi-turn dialogue generation with tier-specific reject_and_refine turns and verbatim review-span citations, (4) automated quality validation (turn structure, action coverage, POI/review references, per-quote fuzzy match >= 0.80 partial ratio), and (5) difficulty annotation (Easy/Medium/Hard/Expert) and stratified train/validation/test split (seed 42).",
"prov:used": [
{
"@type": "prov:Entity",
"rdfs:label": "gpt-5.4-mini",
"description": "Large language model used as both the dialogue generator and the structured-output engine via DSPy signatures. Single-LLM dependence is documented in the dataset's known limitations."
},
{
"@type": "prov:Entity",
"rdfs:label": "DSPy",
"description": "Programmatic-prompting framework providing typed signatures for action transitions, citation pointers, and grounding annotations during generation.",
"url": "https://github.com/stanfordnlp/dspy"
},
{
"@type": "prov:Entity",
"rdfs:label": "Yelp Open Dataset",
"description": "Source of business metadata, review IDs, and review text (used only at generation time; review text is NOT redistributed in the released dataset).",
"url": "https://www.yelp.com/dataset"
}
],
"prov:startedAtTime": "2026-04-23T00:00:00Z",
"prov:endedAtTime": "2026-04-24T23:59:59Z"
},
"rai:dataCollection": "Synthetic dialogues were generated with DSPy and gpt-5.4-mini conditioned on (a) one of 47 persona archetypes, (b) 8 candidate POIs sampled from the Yelp Open Dataset by stratified rating/category, and (c) tier-specific dialogue scenario prompts. The five-stage pipeline is: scenario construction, candidate POI selection, dialogue generation, automated quality validation (turn structure, action coverage, POI/review references, per-quote fuzzy match >= 0.80 partial ratio), and difficulty annotation (Easy/Medium/Hard/Expert).",
"rai:dataAnnotationProtocol": "Per-turn annotations (action label, referenced_poi_ids, referenced_review_ids, acceptable_alternative_poi_ids, citation pointers, discovered_review_ids) are produced jointly with dialogue generation by the LLM under structured DSPy signatures and validated by automated checks. A secondary ALCE-style human evaluation pilot used five trained annotators rating 90 items (30 dialogues x 3 baselines: TF-IDF, RAG-Citation, LLM Zero-Shot) for citation precision, informativeness, and naturalness; Krippendorff's alpha is 0.31-0.58 across labels. Twenty additional dialogues received blind-written expert-author recommendations (POI Jaccard = 0.437, n=215).",
"rai:dataPreprocessingProtocol": "POIs were drawn from the Yelp Open Dataset business and review files (yelp_academic_dataset_business.json, yelp_academic_dataset_review.json), filtered to 8 U.S. cities x 3 POI types x 100 POIs (= 2,400 POIs) with up to 15 reviews per POI (>= 5 minimum). Verbatim review spans cited by the LLM during dialogue generation were detected via fuzzy substring matching (>= 0.90 partial ratio) against the source review and replaced with [Q:N] placeholders. Quote-span metadata stores only (review_id, [start, end]) for exact quotes (~86% of cases) and (review_id, [start, end], edits) for near-verbatim quotes (~11%) where edits use whitelisted punctuation only (no Yelp word content). Splits are stratified by (city, difficulty) with seed 42.",
"rai:dataReleaseMaintenancePlan": "The dataset is hosted on Hugging Face at https://huggingface.co/datasets/taoist/TRACE under CC BY 4.0 for our authored content. Versions follow semver (current: 1.0.0). Issues, bug reports, and dataset corrections are tracked via the HF dataset Discussions page. Maintenance commitment: active corrections for at least 24 months post-publication, with deprecated versions kept available via git revisions. Yelp-content reconstruction is guaranteed against any Yelp Open Dataset snapshot listed in the README compatibility matrix; new snapshots will be tested and added as Yelp publishes them.",
"rai:dataUseCases": "Intended uses: (1) benchmarking conversational recommender systems on the joint axes of accuracy, citation grounding, and recovery; (2) studying the trade-off between LLM-style fluency and verbatim-evidence faithfulness in recommendation; (3) developing rejection-handling and aspect-synthesis techniques for multi-turn dialogue; (4) calibrating automated grounding metrics (CGS, NLI-entailment) against human judgments. Out-of-scope uses: production deployment for tourism advice without independent verification of POI accuracy, freshness, and licensing; commercial use of any Yelp-derived component without separate authorization from Yelp.",
"rai:dataBiases": "Known biases: (1) Geographic - 8 mid-sized U.S. cities only (Indianapolis, Nashville, New Orleans, Philadelphia, Reno, Saint Louis, Tampa, Tucson); not representative of major metros, rural areas, or non-U.S. tourism. (2) Linguistic - U.S. English only; review style and user-preference expression follow Yelp/U.S. conventions. (3) Source - Yelp reviews skew toward businesses with strong online presence and active reviewer communities; long-tail or recently-opened venues are under-represented. (4) Persona - 47 hand-curated personas tilt toward common traveler archetypes; less coverage for accessibility-focused, elderly, or non-leisure travelers. (5) Generator - all dialogues come from a single LLM family (gpt-5.4-mini); stylistic patterns and reasoning biases of that model are inherited. (6) Visibility amplification - selecting top-rated POIs as candidates may reinforce attention to already popular venues.",
"rai:dataLimitations": "Synthetic-dialogue artifact: dialogues are LLM-generated, not transcripts of real users, and may not capture full pragmatic complexity (multi-constraint negotiation, temporal availability, accessibility needs). Single-generator dependence: model-family-sensitive findings (notably rejection recovery) need cross-model replication. Open-set heterogeneity: baselines expose different candidate-access mechanisms; open-set comparisons are system-level rather than pure model quality. Human-eval scale: 90 items with five annotators is a calibration pilot, not a leaderboard. Grounding metric: composite CGS uses a hand-tuned CD/0.05 gate; baseline rankings should be cross-checked under alternative gates. Yelp dependency: the full knowledge base requires user-side reconstruction from a Yelp Open Dataset bundle; review text is not redistributed.",
"rai:dataSocialImpact": "Positive: pushes CRS evaluation toward verifiable, audited recommendations rather than opaque scoring, which is particularly important in high-stakes consumer decisions like travel and hospitality. Risks: (1) over-optimization toward citation-density proxies may encourage excessive verbatim quoting rather than semantically supported reasoning; (2) deployments on Yelp-style data may concentrate attention on already popular venues, contributing to overtourism and neighborhood-level visibility inequities; (3) Yelp reviews encode social, geographic, and demographic biases that propagate to systems trained or evaluated on TRACE; (4) tourism recommendation can be misused to surface or hide specific businesses for commercial gain. Users deploying TRACE-trained systems should audit recommendation distributions for fairness across neighborhoods, price tiers, and minority-owned businesses.",
"rai:personalSensitiveInformation": "The dataset does not contain personal data of real individuals. User personas are synthetic archetypes (e.g., 'budget-conscious solo traveler') and do not correspond to any real user. Yelp reviewer identifiers, user names, and account data are NOT included. POI references identify businesses (Yelp business_id) and review IDs (Yelp review_id) as bare factual identifiers; no reviewer-attributable text is redistributed. Content moderation: dialogues were generated under DSPy structured signatures with a system prompt that excludes personal-identifier fabrication; no manual safety review was performed beyond automated validation."
}