Datasets:

jonaskg
/

open-deliberation

license: cc-by-4.0
language:
  - en
  - zh
  - de
tags:
  - deliberation
  - democracy
  - polis
  - votes
  - habermas-machine
  - collective-dialogue
pretty_name: Open Deliberation
configs:
  - config_name: text_corpus
    data_files: text/deliberation_combined.csv
    default: true
  - config_name: polis_votes
    data_files:
      - split: comments
        path: polis_votes/*/comments.csv
      - split: votes
        path: polis_votes/*/votes.csv
      - split: summary
        path: polis_votes/*/summary.csv

Open Deliberation

Two related releases in one dataset repo:

Config	What it is	Vote-preserving?	Size
`text_corpus`	Pruned multi-source text (Polis comments + Habermas + OpenAI Remesh)	No — only optional agree/disagree tallies in `extra`	~1.8 MB
`polis_votes`	Polis comments + votes + summary per conversation	Yes — participant×comment vote matrix	~lean pack

from datasets import load_dataset

text = load_dataset("jonaskg/open-deliberation", "text_corpus")
votes = load_dataset("jonaskg/open-deliberation", "polis_votes")

Why two configs?

Polis’s power is the vote matrix: each participant voting agree/disagree/pass on each comment, which enables PCA opinion-group maps. A text-only export loses that geometry.
text_corpus is for NLP / scenario / mediation research.
polis_votes is for computational social choice and opinion-map replication.

`text_corpus` columns

source, corpus, record_type, topic, question, text, stance_or_label, agreement_pct, extra

Sources: Polis comments (CC BY 4.0), Habermas Machine (Science 2024), OpenAI Democratic Inputs / Remesh.

`polis_votes` layout

polis_votes/<conversation_id>/
  comments.csv
  votes.csv
  summary.csv

Lean conversations included (vote matrices for March On and the largest Austria Klimarat dumps omitted to keep the release downloadable). Case study index: compdemocracy.org.

Attribution (Polis)

Data was gathered using the Polis software (compdemocracy.org/polis) and is sub-licensed under CC BY 4.0 with Attribution to The Computational Democracy Project.

Notes

Not an endorsement of political positions in the data.
Habermas / Remesh appear only in text_corpus.