#!/usr/bin/env python3 """Generate Croissant 1.1 + minimal RAI metadata for BehaviorBench.""" import hashlib import json from pathlib import Path import pyarrow as pa import pyarrow.parquet as pq ROOT = Path(__file__).resolve().parents[1] BASE = "https://huggingface.co/datasets/Salesforce/BehaviorBench" RAW = BASE + "/resolve/main/" def field_type(t): """Return (Croissant type, repeated) for validator-safe scalar fields. Current mlcroissant record generation cannot reliably cast nested Parquet struct values (including list>). Those columns remain present in the distributed Parquet files but are omitted from RecordSet fields. """ repeated = False if pa.types.is_list(t) or pa.types.is_large_list(t): t = t.value_type repeated = True if pa.types.is_boolean(t): return "sc:Boolean", repeated if pa.types.is_integer(t): return "sc:Integer", repeated if pa.types.is_floating(t) or pa.types.is_decimal(t): return "sc:Float", repeated if pa.types.is_string(t) or pa.types.is_large_string(t): return "sc:Text", repeated return None def sha256_file(path, chunk_size=8 * 1024 * 1024): digest = hashlib.sha256() with path.open("rb") as stream: for chunk in iter(lambda: stream.read(chunk_size), b""): digest.update(chunk) return digest.hexdigest() def canonical_files(): files=[] for top in (ROOT/"dataset", ROOT/"benchmark"): for p in top.rglob("*.parquet"): # Generated evaluation source caches are provenance aids, not record sets. if "evaluation/sources" in p.as_posix(): continue files.append(p) return sorted(files) def main(): distributions=[]; recordsets=[] for p in canonical_files(): rel=p.relative_to(ROOT).as_posix(); fid=rel distributions.append({ "@type":"cr:FileObject", "@id":fid, "name":rel, "contentUrl":RAW+rel, "encodingFormat":"application/x-parquet", "sha256":sha256_file(p), }) pf=pq.ParquetFile(p); schema=pf.schema_arrow rid=BASE+"#records/"+rel.removesuffix(".parquet") fields=[] for f in schema: type_info = field_type(f.type) if type_info is None: continue data_type, repeated = type_info field = { "@type":"cr:Field", "@id":rid+"/"+f.name, "name":f.name, "dataType":data_type, "source":{"fileObject":fid,"extract":{"column":f.name}}, } if repeated: field["repeated"] = True fields.append(field) recordsets.append({ "@type":"cr:RecordSet", "@id":rid, "name":rel, "description":f"Rows from {rel} ({pf.metadata.num_rows:,} records).", "field":fields, }) doc={ "@context":{ "@language":"en", "@vocab":"https://schema.org/", "sc":"https://schema.org/", "cr":"http://mlcommons.org/croissant/", "rai":"http://mlcommons.org/croissant/RAI/", "dct":"http://purl.org/dc/terms/", "prov":"http://www.w3.org/ns/prov#", "arrayShape":"cr:arrayShape", "citeAs":"cr:citeAs", "column":"cr:column", "conformsTo":"dct:conformsTo", "containedIn":"cr:containedIn", "data":{"@id":"cr:data","@type":"@json"}, "dataType":{"@id":"cr:dataType","@type":"@vocab"}, "description":{"@container":"@language"}, "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", "isArray":"cr:isArray", "isLiveDataset":"cr:isLiveDataset", "jsonPath":"cr:jsonPath", "key":"cr:key", "md5":"cr:md5", "name":{"@container":"@language"}, "parentField":"cr:parentField", "path":"cr:path", "recordSet":"cr:recordSet", "references":"cr:references", "regex":"cr:regex", "repeated":"cr:repeated", "replace":"cr:replace", "samplingRate":"cr:samplingRate", "separator":"cr:separator", "source":"cr:source", "subField":"cr:subField", "transform":"cr:transform" }, "@type":"sc:Dataset", "name":"BehaviorBench", "alternateName":"BehaviorBench: Modeling Real-World User Decisions from Behavioral Traces", "description":"A benchmark derived from public Polymarket and Polygon behavioral traces for wallet-level Belief and Trade prediction under Direct, Profile, and Retrieval interfaces.", "url":BASE, "license":"https://www.apache.org/licenses/LICENSE-2.0", "conformsTo":"http://mlcommons.org/croissant/1.1", "citeAs":"BehaviorBench: Modeling Real-World User Decisions from Behavioral Traces (NeurIPS 2026 E&D submission).", "version":"1.0.0", "datePublished":"2026", "inLanguage":"en", "keywords":["behavioral modeling","user modeling","prediction markets","Polymarket","personalization","benchmark"], "isAccessibleForFree":True, "distribution":distributions, "recordSet":recordsets, "rai:dataLimitations":"Observable wallet behavior is an imperfect proxy for internal belief or preference. The sample represents active Polymarket wallets rather than the general population. Belief confidence is concentrated at 0 and 1; ties do not uniquely identify direction. Trade values use estimated mark-price conversion. Generalization beyond this platform and block range is not established.", "rai:dataBiases":"Activity filtering favors active wallets; automated strategies, market makers, hedgers, and arbitrageurs may be overrepresented. Market topics and participation are time-dependent. Generated profiles inherit biases from their generator models.", "rai:personalSensitiveInformation":"The data contain public pseudonymous blockchain wallet addresses and behavioral traces that may support sensitive inferences about financial interests, risk preferences, politics, or sports. No verified real-world identities or demographic attributes are included. Deanonymization and consequential use are not intended.", "rai:dataUseCases":"Intended for research evaluation of personalized behavioral prediction, sequence modeling, profile representations, and retrieval augmentation. Not intended for identity attribution, financial advice, surveillance, credit/employment decisions, law enforcement, or targeting individuals.", "rai:dataSocialImpact":"Potential benefits include reproducible evaluation of personalized systems. Risks include surveillance, profiling, deanonymization, manipulation, and overclaiming psychological inference from financial behavior. Mitigations are documented in docs/RESPONSIBLE_AI.md.", "rai:hasSyntheticData":True, "rai:dataCollection":"Derived from public Polymarket and Polygon on-chain records covering released block heights approximately 30,000,825 through 80,000,000. The release contains processed source tables rather than a complete raw chain archive.", "rai:dataReleaseMaintenance":"Version material corrections in CHANGELOG.md; regenerate checksums and Croissant metadata for each release.", "rai:machineAnnotationTools":"Belief and Trade labels are deterministic transformations of holdings and transactions. Profile interface fields are language-model-generated summaries and retain generator-model metadata where available.", "prov:wasDerivedFrom":[{"@type":"sc:Dataset","name":"Public Polymarket and Polygon records","url":"https://polymarket.com/"}], "prov:wasGeneratedBy":[ {"@type":"sc:CreateAction","name":"Belief construction","description":"Aggregate wallet-market holdings into stance and normalized position imbalance, then create fixed per-wallet temporal splits."}, {"@type":"sc:CreateAction","name":"Trade construction","description":"Order transactions by block/log index, reconstruct position changes, derive increase/decrease/close labels and estimated USDC amount, then create fixed per-wallet temporal splits."}, {"@type":"sc:CreateAction","name":"Benchmark interfaces","description":"Render Direct prompts, model-generated Profile representations, and Retrieval contexts from support wallets excluded from the experiment set."} ] } (ROOT/"croissant.json").write_text(json.dumps(doc,ensure_ascii=False,indent=2)+"\n") print(f"Wrote {len(distributions)} distributions and {len(recordsets)} record sets") if __name__=="__main__": main()