CCDD · Chinese Competitive Debate Dataset

Chinese Competitive Debate: A Dataset and Benchmark with Professional Adjudication

CCDD links transcripts of real Chinese competitive debate to the original adjudication produced under a predefined rubric: individual judge scores for every evaluated stage and side, match decisions, best-debater ballots and commentary, with linked appeal reassessments. Three benchmark tasks — winner prediction, stage-score prediction and best-debater prediction — measure agreement with professional adjudication at the match, stage and speaker levels.

Dataset repositoryData documentationBenchmark
01 · At a glance
Matches
148

Retained from 182 source matches after filtering incomplete or unresolved scoring records.

Stage–side instances
2,690

One scheduled stage × one evaluated side, each with three original judge scores (1–10).

Turns
20,542

Single-side contributions extracted from interactive stages; linked to parent-stage assessments.

Panel adjudications
168

148 original panels plus 20 appeal panels; every panel is three independent judges.

02 · What one record links

Match

Motion, format and the ordered transcript; three judge decisions consolidated from impression, stage-based and decisive votes; the panel outcome; nine best-debater votes; recorded commentary.

Stage

Scheduled segments with a procedural type. A speech stage yields one scored side; questioning, cross-examination, head-to-head and free debate yield two, each with its own three ratings.

Turn

One continuous contribution from one side inside an interactive stage. Turns keep their order and parent stage but carry no independent quality score.

03 · Corpus and judge agreement

Corpus statistics

QuantityCount
Source / excluded matches182 / 34
Retained matches148
  4v4 without head-to-head71
  4v4 with head-to-head45
  2v232
Stage–side instances2,690
Turns in interactive stages20,542
Original panel adjudications148
Appeal panel adjudications20
Total panel adjudications168

Judge agreement and appeals

StatisticValue
Unanimous panels82 / 148
Split panels66 / 148
Pairwise decision agreement70.3%
Squared deviation from panel mean (stage scores)0.759
Score-band agreement (all / exactly two / none)52.8 / 44.3 / 2.8%
ICC(1,3): stage–side scores0.523
ICC(1,3): paired affirmative − negative differences0.630
Appeal reversals7 / 20 (35%)
04 · Zero-shot benchmark results

Task 1 · Winner prediction

Accuracy over 148 matches
opus-50.662
deepseek-v4-pro0.615
deepseek-v4-flash (Reasoning)0.601
gpt-5.6-sol0.601
gemini-3.5-flash-lite0.595
sonnet-50.595
deepseek-v4-flash0.588
haiku-4.50.588
gpt-5.6-luna0.561
Always affirmative0.507
Random0.500

Task 2 · Stage-score prediction

Pearson r with the three-judge mean, 2,690 instances
deepseek-v4-flash0.250
deepseek-v4-flash (Reasoning)0.241
deepseek-v4-pro0.231
gemini-3.5-flash-lite0.184
gpt-5.6-luna0.232
Random0.000
Structure-aware random0.031

Task 3 · Best-debater prediction

Selection accuracy over 148 matches
gpt-5.6-sol0.568
deepseek-v4-pro0.554
opus-50.520
deepseek-v4-flash (Reasoning)0.439
deepseek-v4-flash0.432
gpt-5.6-luna0.432
sonnet-50.419
haiku-4.50.399
gemini-3.5-flash-lite0.345
Random0.183

Zero-shot, temperature 0 where supported, structured outputs. Dashed outlines are references. Full tables, metric definitions, structural and human references are on the benchmark page.

05 · Documentation
01
Construction
Rubric, competition-time adjudication, processing, alignment, filtering
02
Data documentation
Hierarchy, schema, package layout, release components
03
Benchmark
Tasks, metrics, protocol, results, references
04
Judges and rubric
Judge pool requirements and the full 1–10 rubric
05
License and terms
What you may do with the data
06
Privacy
De-identification, consent clause, withdrawal
07
Paper
Abstract, contributions, comparison with prior resources
06 · Download and citation

Package

CCDD v1.0: full/ (148 match records + task files), multi_panel/ (20 appealed matches, both panels), cross_format/ (4v4 and 2v2 subsets). About 2.5 GB of JSON/JSONL.

Access

The dataset is available at anonymous.4open.science/r/CCDD. The six largest exchange-level files are stored as parts under 100 MB; the README explains how to concatenate them.

Cite

@misc{yang2026ccdd,
  title  = {Chinese Competitive Debate: A Dataset and
            Benchmark with Professional Adjudication},
  author = {Yang, Zongrui and Li, Haoyuan and Wang, Zhongsheng and Zeng, Zhirui and Han, Pengqian and Zhou, Yi and Wang, Yuting and Liu, Jiamou},
  year   = {2026},
  eprint = {2609.21637},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  url    = {https://arxiv.org/abs/2609.21637}
}

07 · Acknowledgements

We thank all friends in the debating community who offered advice, feedback and help with data processing for this benchmark.

We thank Mr. Xi Rui (席瑞), Mr. Wan Huaming (万华明), Ms. Wu Xinning (仵心凝) and other senior figures for their insight and ideas; and Beiming Debate Club (北冥辩论俱乐部), Adventurers' Tavern Debate Club (冒险者酒馆辩论俱乐部), the New Zealand Chinese Debating Association (新西兰华语辩论协会) and other organisations for their support and advice throughout data review, processing, tournament organisation and evaluation design.

Above all, we thank every expert judge who took part. Your long-accumulated debating practice and adjudication experience, and the care you brought to judging each match, form the most important professional foundation of this dataset and benchmark, and its most valuable asset.