DeliChess

DeliChess is an English multi-party dialogue dataset for studying group deliberation in chess puzzle solving. Participants first selected moves independently, deliberated synchronously in groups, and then submitted revised answers. The corpus contains 107 dialogues, 7,667 chat utterances, and 321 dialogue--puzzle cases. Its event stream also records the puzzle options and paired pre-/post-deliberation choices.

Every genuine chat utterance has human annotations for communicative function, epistemic stance, and deliberative-agent usefulness. This release additionally includes independent Gemini 3.5 Flash predictions for the same three dimensions on all 7,667 utterances. The Gemini predictions are an auxiliary comparison layer, not replacements for the human annotations.

Paper: DeliChess: A Multi-party Dialogue Dataset for Deliberation in Chess Puzzle Solving

Dataset configurations

• utterances (default): one row per genuine chat utterance, with clean text, dialogue-local speaker identifiers, human labels, Gemini predictions, and agreement flags.
• events: the complete 12,596-row ordered task record. It includes chat, puzzle specifications, independent and revised submissions, and task-state events. Annotation fields are blank for non-chat events.
• dialogues: one privacy-safe summary row per dialogue.

All data are provided as UTF-8 CSV. Hugging Face calls the sole split train for loading purposes, but the release does not prescribe a train/test split. Researchers should split by dialogue_id, not by utterance.

See USAGE.md for loading examples and docs/ANNOTATION_GUIDELINES.md for the complete annotation definitions.

Annotation schema

Communicative function

PROPOSE_ANSWER, PROVIDE_REASONING, EXPLORE_ALTERNATIVES, REQUEST_REASONING, EVALUATE_CRITIQUE, AGREEMENT_ALIGNMENT, COORDINATE_DECISION, SOCIAL_MODERATION, and SOCIAL_OFFTASK.

Epistemic stance

NO_TASK_STANCE, HEDGED, and NO_HEDGED. The last is the historical name for an unhedged task stance.

Deliberative-agent usefulness

HELPFUL, NEUTRAL, NOT_APPLICABLE, and HARMFUL. This dimension asks whether a conversational move would be useful for an agent supporting group deliberation; it is not a generic sentiment or quality label.

Data fields

The utterances configuration contains:

• anonymous dialogue, utterance, and event IDs;
• within-dialogue utterance and event positions;
• dialogue-local speaker ID and study-assigned animal alias;
• cleaned utterance text and time elapsed from the dialogue's first event;
• three human_* annotation columns;
• three gemini_* prediction columns;
• per-dimension and all-three agreement flags; and
• an optional, privacy-screened annotator note.

The events configuration additionally contains the event type, original privacy-screened event payload, task-state fields, and blank annotation fields for non-chat rows. WASON_INITIAL stores the three puzzle specifications and candidate moves; GAME_SUBMIT and WASON_SUBMIT record independent and revised selections, respectively.

Dataset creation

The task pool contains 24 chess puzzles: eight endgame, eight positional, and eight tactical. Each dialogue uses one puzzle from each category and offers five candidate moves, including the Stockfish-best move and four plausible distractors. Participants first answer independently, then deliberate, and finally revise their choices.

The corpus annotations were collected from trained proficient English speakers recruited through Prolific. The complete guidelines are included in this repository. The study protocol received institutional ethics approval and participants gave informed consent.

For the auxiliary model layer, Gemini received the complete dialogue context and the same canonical guidelines but never the human labels. Inference used Gemini 3.5 Flash at temperature 0. Detailed configuration and agreement results are documented in docs/GEMINI_ANNOTATION.md.

Anonymisation and sensitive content

The release contains no Prolific IDs, assignment/session IDs, source message or user IDs, absolute timestamps, private paths, contact details, or reverse mapping. Speakers are reindexed independently within each dialogue. The animal names in speaker_name are aliases assigned by the study, not real names. Incidental personal names and demographic disclosures found during screening were masked.

The dialogues are task-focused. Manual screening found no hate speech, threats, or targeted abuse, although a small number of mild expletives remain as part of the natural dialogue record.

Intended uses and limitations

DeliChess supports research on multi-party dialogue, collaborative reasoning, dialogue acts, epistemic stance, deliberation outcomes, and deliberation-support agents. It should not be used to identify participants or infer sensitive attributes.

The data come from a controlled, text-only chess task and may not generalize to face-to-face deliberation, other domains, or other participant populations. Most main-corpus utterances have one human annotation, so the release does not provide independent duplicate labels for every dialogue. Utterances are nested within dialogues and should not be treated as independent observations.

Gemini agreement with the human annotations is uneven across dimensions and labels. These predictions are model outputs rather than adjudicated annotation, and downstream users should keep their provenance explicit.

Licensing

The dataset is released under the MIT License. See LICENSE.

Citation

@misc{zhu2026delichessmultipartydialoguedataset,
      title={DeliChess: A Multi-party Dialogue Dataset for Deliberation in Chess Puzzle Solving}, 
      author={Xiaochen Zhu and Georgi Karadzhov and Tom Stafford and Andreas Vlachos},
      year={2026},
      eprint={2606.04987},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.04987}, 
}