Documentation · 02 of 07

Data documentation

Text is organised as Match → Stage → Turn. Stages link to side-specific assessments, speakers link across their contributions, and match identifiers associate original and appeal adjudications. Human records and derived aggregates are kept distinguishable.

01 · Entities, links and provenance
Entity or field groupValue type and linkageMeaning and scope
MatchIdentifier; text; format categoryOne unique debate with its motion, format and ordered transcript. match_id links every derived row to its source debate.
StageOrdered segment; name and typeOne scheduled segment. Separate speeches remain separate stages; a shared interactive segment remains one stage.
TurnOrdered text; parent stage and sideOne continuous single-side contribution from an interactive stage (legacy field name exchange). No independent quality score.
SpeakerWithin-match identity and side; positionA participant whose contributions can span stages, identified by role (e.g. affirmative speaker 3). Distinct from the evaluated side of a stage.
Corrected text unittext; text_rawSentence-level correction output alongside the raw ASR line; not necessarily a complete turn.
Stage–side assessmentStage and side; three ratingsOne valid scored target with original ratings on the 1–10 scale. A shared stage can support two side-specific targets.
Outcome componentsPer-judge vote allocationsImpression, stage-based and decisive votes used in within-judge consolidation; distinct from best-debater ballots.
Judge decisionSide-valued; judge and matchOne consolidated outcome decision per judge, derived from that judge's components.
Best-debater ballotsPer-speaker counts; judge and matchThree votes per judge, nine across the original panel. Individual allocations and panel totals have distinct provenance.
Judge commentaryText; match and judgeRecorded human commentary in processed transcription; distinct from generated summaries.
Appeal adjudicationAdditional judgments; same matchAvailable for the 20 reassessed matches; preserved separately from original judgments.
Derived targetsMeans, vote totals, outcomesAggregates computed from identified human records. The benchmark uses original adjudication.
02 · Format coverage and units
Format variantMatchesSingle-side scored stages / matchPaired-side scored stages / matchInstances / matchTotal instances
4v4 without head-to-head7185181,278
4v4 with head-to-head458620900
2v2328416512
Total148———2,690

Single-side stages contribute one instance and paired stages two. Optional 2v2 surprise challenges are unscored. Best-debater selection involves eight candidates in 4v4 and four in 2v2.

03 · Package layout (v1.0)
CCDD-v1.0/
  README.md
  full/            matches/  148 JSON, one per match: transcript units (text, text_raw),
                              per-judge stage scores, votes, best-debater votes, commentary
                   tasks/    match_outcome, stage_speech / question / answer / duel / free_debate,
                              item_question / answer / free_debate, and 6 appeal_* files
  multi_panel/     20 appealed matches: transcript, first_panel, second_panel;
                   paired_match_outcome.jsonl (20), paired_stage_scores.jsonl (261)
  cross_format/    4v4/ (116 matches) and 2v2/ (32 matches), same structure as full/, no appeals

Task rows share match_id, cup, format, motion, outcome. match_outcome adds the full context_text, teams and best_debater_votes. stage_* rows hold one stage × one scored side with text, cumulative context_text, the three judges' commentary and stage_scores. item_* rows hold one turn pair (question, answer, target_text, exchange_id) and inherit the stage label (label_granularity = stage_shared). Appeal files carry panel: appeal.

Row counts: stage level 2,690 main (4v4-A 18, 4v4-B 20 and 2v2 16 per match) plus 362 appeal rows; exchange level 20,542. Judge signals (one text × one judge): 504 match-level, 9,152 stage-level, 61,555 exchange-level.

04 · Public components
Public componentContents
Debate textDe-identified transcripts and their structural organisation
Individual judge scoresSide-specific stage assessments with adjudication provenance
Match results and awardsOriginal-panel winner, best-debater votes, underlying judge decisions and ballots
Judge commentaryText of recorded adjudication explanations
Appeal recordsReassessments and outcome changes linked to original matches
Generated summariesSeparately identified model-generated commentary summaries, distributed apart from the human records
← ConstructionBenchmark →