| { |
| "@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", |
| "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", |
| "samplingRate": "cr:samplingRate", |
| "sc": "https://schema.org/", |
| "separator": "cr:separator", |
| "source": "cr:source", |
| "subField": "cr:subField", |
| "transform": "cr:transform" |
| }, |
| "@type": "sc:Dataset", |
| "name": "Debate Collapse/Correction Evaluation Artifact", |
| "description": "Validated Croissant/RAI metadata for the collapse/correction evaluation protocol bundle used in a homogeneous-panel LLM debate audit.", |
| "conformsTo": "http://mlcommons.org/croissant/1.0", |
| "citeAs": "Xin Li, Mengbing Liu, and Chau Yuen. Measuring Collapse and Correction in Homogeneous-Panel LLM Debate. NeurIPS 2026 Evaluations and Datasets Track. DebateLedger artifact version 1.0.0.", |
| "dateCreated": "2026-05-02", |
| "datePublished": "2026-09-28", |
| "inLanguage": "en", |
| "keywords": [ |
| "multi-agent debate", |
| "evaluation", |
| "collapse", |
| "correction", |
| "MMLU-Pro", |
| "responsible AI" |
| ], |
| "license": "https://creativecommons.org/licenses/by/4.0/", |
| "rai:dataAnnotationProtocol": "No human labeling. Collapse/correction labels are deterministic functions of parsed MCQ answers and answer keys.", |
| "rai:dataBiases": "Primary evidence is MMLU-Pro; homogeneous same-model three-agent scaffold; realized G=7 family cohort with Qwen-heavy model rows.", |
| "rai:dataCollection": "Public benchmark MCQ questions were processed by LLM agents under fixed probe and debate prompts. No human-subjects or personal data were collected.", |
| "rai:dataCollectionMissingData": "Closed-API snapshot metadata are partial for older sealed lanes; full transcripts and the R1 feature matrix are gated for dual-use reasons.", |
| "rai:dataCollectionRawData": "Upstream public MCQ benchmark questions are not owned by this artifact; derived aggregate tables and trace summaries are released subject to upstream benchmark terms.", |
| "rai:dataCollectionType": "machine-generated traces from public benchmark questions", |
| "rai:dataLimitations": "Not a benchmark-general law; not a calibrated per-question oracle; not a deployable controller; gated tier needed for convince-wrong templates and full transcripts.", |
| "rai:dataPreprocessingProtocol": "Rule-based MCQ answer extraction; majority voting after filtering unparsed answers; alphabetic tie-breaking; deterministic collapse/correction/onset labels.", |
| "rai:dataReleaseMaintenancePlan": "Versioned public release (alpha-tot-v1.0): open data on Hugging Face, code on GitHub, and a gated Hugging Face dataset for full transcripts and convince-wrong templates. Corrections and provider or parser changes are released as new minor versions.", |
| "rai:dataSocialImpact": "Helps audit debate interventions that may reduce collapses while discarding corrections; gated release mitigates misuse as a misleading-answer corpus.", |
| "rai:dataUseCases": "Zero-API recomputation of aggregate tables; applying the transition-table and signed-utility protocol to new model-scaffold pairs; documenting gated release boundaries; public artifact hosting on Hugging Face with code on GitHub.", |
| "rai:hasSyntheticData": true, |
| "rai:machineAnnotationTools": "Rule-based MCQ answer parser included in the artifact; no LLM judge step is used.", |
| "rai:personalSensitiveInformation": "none", |
| "url": "https://huggingface.co/datasets/XINLI1997/DebateLedger", |
| "version": "1.0.0", |
| "distribution": [ |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_README_REPRO_md", |
| "name": "README_REPRO.md", |
| "contentUrl": "README_REPRO.md", |
| "encodingFormat": "text/markdown", |
| "sha256": "490012234e940b5942ca086b8e31e2f78801c52953a87d3db2bad2ae7bd8709a" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_requirements_artifact_txt", |
| "name": "requirements-artifact.txt", |
| "contentUrl": "requirements-artifact.txt", |
| "encodingFormat": "text/plain", |
| "sha256": "a93bd89c724fa0d4b79f20dfb62bbf356407549157c4e98b0bcf6b970171a14d" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_DEBATE_EVAL_ARTIFACT_DATASHEET_md", |
| "name": "DEBATE_EVAL_ARTIFACT_DATASHEET.md", |
| "contentUrl": "abc_exp/results/DEBATE_EVAL_ARTIFACT_DATASHEET.md", |
| "encodingFormat": "text/markdown", |
| "sha256": "2a7e2e4f882bcd16a4ee7134888699963c018e57f23ececc08ba705332bce4bb" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_debate_eval_reuse_card_schema_json", |
| "name": "debate_eval_reuse_card.schema.json", |
| "contentUrl": "abc_exp/results/debate_eval_reuse_card.schema.json", |
| "encodingFormat": "application/json", |
| "sha256": "4d6c028c58cd501c8fac851b5cdb07ca58c73ee81b34b63c0ffe5a7e8ed8015f" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_cohort_n14_headline_json", |
| "name": "cohort_n14_headline.json", |
| "contentUrl": "abc_exp/results/cohort_n14_headline.json", |
| "encodingFormat": "application/json", |
| "sha256": "83568df98c0bc013e55c8f45a8eed1ea95ac42a67a4588d37b89db005e07cd9d" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_family_level_headline_aggregation_json", |
| "name": "family_level_headline_aggregation.json", |
| "contentUrl": "abc_exp/results/family_level_headline_aggregation.json", |
| "encodingFormat": "application/json", |
| "sha256": "11fc1b8de06a396b24b39119bb0dac1d79a1ee05cb3037ed9f63d632f467e049" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_conditional_collapse_table_json", |
| "name": "conditional_collapse_table.json", |
| "contentUrl": "abc_exp/results/conditional_collapse_table.json", |
| "encodingFormat": "application/json", |
| "sha256": "2df5c2ae2f8047c2d5706fe82339229774d4614c0583351042034f07ccaa5eba" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_reviewer_hardening_analyses_json", |
| "name": "reviewer_hardening_analyses.json", |
| "contentUrl": "abc_exp/results/reviewer_hardening_analyses.json", |
| "encodingFormat": "application/json", |
| "sha256": "5fc36d977086d7e3b2a9b2ccaf408670d8dbd520bf7ff4a7691b479bc2feac03" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_N14_REVIEWER_HARDENING_ARTIFACTS_md", |
| "name": "N14_REVIEWER_HARDENING_ARTIFACTS.md", |
| "contentUrl": "abc_exp/results/N14_REVIEWER_HARDENING_ARTIFACTS.md", |
| "encodingFormat": "text/markdown", |
| "sha256": "14d5d3a1e220330b46d1572e1a808e3b09a01cd3295437e505e38c3a9b148775" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_n14_reviewer_hardening_artifacts_json", |
| "name": "n14_reviewer_hardening_artifacts.json", |
| "contentUrl": "abc_exp/results/n14_reviewer_hardening_artifacts.json", |
| "encodingFormat": "application/json", |
| "sha256": "86f7dcd2a28e3a586ac041a3ab09e45514f45392ca10439cd81a96964a877273" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_ROUND4_CAMERA_READY_HARDENING_md", |
| "name": "ROUND4_CAMERA_READY_HARDENING.md", |
| "contentUrl": "abc_exp/results/ROUND4_CAMERA_READY_HARDENING.md", |
| "encodingFormat": "text/markdown", |
| "sha256": "8f14b8dfbb908c868b8c1894688d1d16c355f40ac395019b2934359c3877c611" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_ROUND5_API_CROSSBENCH_HARDENING_md", |
| "name": "ROUND5_API_CROSSBENCH_HARDENING.md", |
| "contentUrl": "abc_exp/results/ROUND5_API_CROSSBENCH_HARDENING.md", |
| "encodingFormat": "text/markdown", |
| "sha256": "cda64761dab8e7c2b056187213b784c8891c5d11619e284f9b2ee5630357c7a9" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_round5_api_crossbench_hardening_json", |
| "name": "round5_api_crossbench_hardening.json", |
| "contentUrl": "abc_exp/results/round5_api_crossbench_hardening.json", |
| "encodingFormat": "application/json", |
| "sha256": "a117f6e0e01e8a5a080f9dfd13f5570da7f4893d93e1376b3cfc69c9cf6e693d" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_PARSER_CELL_STABILITY_AUDIT_md", |
| "name": "PARSER_CELL_STABILITY_AUDIT.md", |
| "contentUrl": "abc_exp/results/PARSER_CELL_STABILITY_AUDIT.md", |
| "encodingFormat": "text/markdown", |
| "sha256": "5549c6e54e81772a22e74f87a2fa1364a6e9e1241b51d0cd040b7f7a59c2d2aa" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_parser_cell_stability_audit_json", |
| "name": "parser_cell_stability_audit.json", |
| "contentUrl": "abc_exp/results/parser_cell_stability_audit.json", |
| "encodingFormat": "application/json", |
| "sha256": "4f05af44a6d9f76b33c1a8048f23b49d33ee9408f2d5ff9f39136a9e72e9ab56" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_correction_preserving_policy_replay_json", |
| "name": "correction_preserving_policy_replay.json", |
| "contentUrl": "abc_exp/results/correction_preserving_policy_replay.json", |
| "encodingFormat": "application/json", |
| "sha256": "d966b0587b5eb8938ef6aff13fe97fd1f67a19215b2de13b4e3da84bf8c6f09f" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_CORRECTION_PRESERVING_POLICY_REPLAY_md", |
| "name": "CORRECTION_PRESERVING_POLICY_REPLAY.md", |
| "contentUrl": "abc_exp/results/CORRECTION_PRESERVING_POLICY_REPLAY.md", |
| "encodingFormat": "text/markdown", |
| "sha256": "4085b3c32a02c3beefc8cbb87a45757766f1074bac3060a3a03fd83e12f218db" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_block0_alpha_jsonl", |
| "name": "block0_alpha.jsonl", |
| "contentUrl": "abc_exp/results/block0_alpha.jsonl", |
| "encodingFormat": "application/jsonlines", |
| "sha256": "6d5ad54efe4eeb055d7d220dd31f753edceefcd220dab52f7a886adb602092a9" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_block0_debates_jsonl", |
| "name": "block0_debates.jsonl", |
| "contentUrl": "abc_exp/results/block0_debates.jsonl", |
| "encodingFormat": "application/jsonlines", |
| "sha256": "c3c05b9b3ca388cdabc802109b6d1a166e56d10daf65a6bc8f0db64a7fa30274" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_block1_abc_vs_baselines_jsonl", |
| "name": "block1_abc_vs_baselines.jsonl", |
| "contentUrl": "abc_exp/results/block1_abc_vs_baselines.jsonl", |
| "encodingFormat": "application/jsonlines", |
| "sha256": "7cba65bb6ab84ba8c870652ba99c968f4ebe138e9cdb2028214a7a85f708092c" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "file_abc_exp_results_block3_multi_benchmark_jsonl", |
| "name": "block3_multi_benchmark.jsonl", |
| "contentUrl": "abc_exp/results/block3_multi_benchmark.jsonl", |
| "encodingFormat": "application/jsonlines", |
| "sha256": "72a7397af3a6469e64cc2412b58faca14c9dbd3366c8d3114b2d2093d82f9e5e" |
| } |
| ], |
| "recordSet": [ |
| { |
| "@type": "cr:RecordSet", |
| "@id": "artifact_file_inventory", |
| "name": "artifact_file_inventory", |
| "description": "Core files in the public DebateLedger release that define the reusable collapse/correction evaluation, zero-API rebuild, and signed-replay ledgers.", |
| "key": { |
| "@id": "artifact_path" |
| }, |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "artifact_path", |
| "name": "artifact_path", |
| "description": "Repository-relative path for a core artifact file.", |
| "dataType": "sc:Text" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "artifact_role", |
| "name": "artifact_role", |
| "description": "Role of the file in reproducing or reusing the evaluation protocol.", |
| "dataType": "sc:Text" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "encoding_format", |
| "name": "encoding_format", |
| "description": "Media type recorded in the Croissant distribution entry.", |
| "dataType": "sc:Text" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "sha256", |
| "name": "sha256", |
| "description": "SHA-256 digest of the checked-in file.", |
| "dataType": "sc:Text" |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "release_tier", |
| "name": "release_tier", |
| "description": "Release tier of the file in the public release.", |
| "dataType": "sc:Text" |
| } |
| ], |
| "data": [ |
| { |
| "artifact_path": "README_REPRO.md", |
| "artifact_role": "zero-api rebuild card", |
| "encoding_format": "text/markdown", |
| "sha256": "490012234e940b5942ca086b8e31e2f78801c52953a87d3db2bad2ae7bd8709a", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "requirements-artifact.txt", |
| "artifact_role": "Python dependency manifest", |
| "encoding_format": "text/plain", |
| "sha256": "a93bd89c724fa0d4b79f20dfb62bbf356407549157c4e98b0bcf6b970171a14d", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/DEBATE_EVAL_ARTIFACT_DATASHEET.md", |
| "artifact_role": "artifact datasheet", |
| "encoding_format": "text/markdown", |
| "sha256": "2a7e2e4f882bcd16a4ee7134888699963c018e57f23ececc08ba705332bce4bb", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/debate_eval_reuse_card.schema.json", |
| "artifact_role": "new-row reuse schema", |
| "encoding_format": "application/json", |
| "sha256": "4d6c028c58cd501c8fac851b5cdb07ca58c73ee81b34b63c0ffe5a7e8ed8015f", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/cohort_n14_headline.json", |
| "artifact_role": "N14 headline rows", |
| "encoding_format": "application/json", |
| "sha256": "83568df98c0bc013e55c8f45a8eed1ea95ac42a67a4588d37b89db005e07cd9d", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/family_level_headline_aggregation.json", |
| "artifact_role": "family-level primary aggregation", |
| "encoding_format": "application/json", |
| "sha256": "11fc1b8de06a396b24b39119bb0dac1d79a1ee05cb3037ed9f63d632f467e049", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/conditional_collapse_table.json", |
| "artifact_role": "collapse/correction transition table", |
| "encoding_format": "application/json", |
| "sha256": "2df5c2ae2f8047c2d5706fe82339229774d4614c0583351042034f07ccaa5eba", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/reviewer_hardening_analyses.json", |
| "artifact_role": "review audit diagnostics", |
| "encoding_format": "application/json", |
| "sha256": "5fc36d977086d7e3b2a9b2ccaf408670d8dbd520bf7ff4a7691b479bc2feac03", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/N14_REVIEWER_HARDENING_ARTIFACTS.md", |
| "artifact_role": "N14 diagnostic summary", |
| "encoding_format": "text/markdown", |
| "sha256": "14d5d3a1e220330b46d1572e1a808e3b09a01cd3295437e505e38c3a9b148775", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/n14_reviewer_hardening_artifacts.json", |
| "artifact_role": "N14 diagnostic data", |
| "encoding_format": "application/json", |
| "sha256": "86f7dcd2a28e3a586ac041a3ab09e45514f45392ca10439cd81a96964a877273", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/ROUND4_CAMERA_READY_HARDENING.md", |
| "artifact_role": "revision audit summary", |
| "encoding_format": "text/markdown", |
| "sha256": "8f14b8dfbb908c868b8c1894688d1d16c355f40ac395019b2934359c3877c611", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/ROUND5_API_CROSSBENCH_HARDENING.md", |
| "artifact_role": "cross-benchmark stress summary", |
| "encoding_format": "text/markdown", |
| "sha256": "cda64761dab8e7c2b056187213b784c8891c5d11619e284f9b2ee5630357c7a9", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/round5_api_crossbench_hardening.json", |
| "artifact_role": "cross-benchmark stress data", |
| "encoding_format": "application/json", |
| "sha256": "a117f6e0e01e8a5a080f9dfd13f5570da7f4893d93e1376b3cfc69c9cf6e693d", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/PARSER_CELL_STABILITY_AUDIT.md", |
| "artifact_role": "parser stability summary", |
| "encoding_format": "text/markdown", |
| "sha256": "5549c6e54e81772a22e74f87a2fa1364a6e9e1241b51d0cd040b7f7a59c2d2aa", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/parser_cell_stability_audit.json", |
| "artifact_role": "parser stability data", |
| "encoding_format": "application/json", |
| "sha256": "4f05af44a6d9f76b33c1a8048f23b49d33ee9408f2d5ff9f39136a9e72e9ab56", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/correction_preserving_policy_replay.json", |
| "artifact_role": "signed replay ledger", |
| "encoding_format": "application/json", |
| "sha256": "d966b0587b5eb8938ef6aff13fe97fd1f67a19215b2de13b4e3da84bf8c6f09f", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/CORRECTION_PRESERVING_POLICY_REPLAY.md", |
| "artifact_role": "signed replay summary", |
| "encoding_format": "text/markdown", |
| "sha256": "4085b3c32a02c3beefc8cbb87a45757766f1074bac3060a3a03fd83e12f218db", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/block0_alpha.jsonl", |
| "artifact_role": "probe trace sample", |
| "encoding_format": "application/jsonlines", |
| "sha256": "6d5ad54efe4eeb055d7d220dd31f753edceefcd220dab52f7a886adb602092a9", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/block0_debates.jsonl", |
| "artifact_role": "debate trace sample", |
| "encoding_format": "application/jsonlines", |
| "sha256": "c3c05b9b3ca388cdabc802109b6d1a166e56d10daf65a6bc8f0db64a7fa30274", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/block1_abc_vs_baselines.jsonl", |
| "artifact_role": "baseline comparison trace rows", |
| "encoding_format": "application/jsonlines", |
| "sha256": "7cba65bb6ab84ba8c870652ba99c968f4ebe138e9cdb2028214a7a85f708092c", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| }, |
| { |
| "artifact_path": "abc_exp/results/block3_multi_benchmark.jsonl", |
| "artifact_role": "boundary benchmark trace rows", |
| "encoding_format": "application/jsonlines", |
| "sha256": "72a7397af3a6469e64cc2412b58faca14c9dbd3366c8d3114b2d2093d82f9e5e", |
| "release_tier": "open; full transcripts and convince-wrong templates are in the gated dataset" |
| } |
| ] |
| } |
| ] |
| } |
|
|
