language:

• da
  license: mit
  task_categories:
• text-generation
• conversational
  tags:
• danish
• politics
• personas
• steering-vectors
• fine-tuning
• sft
  pretty_name: Danish Political Personas
  size_categories:
• 1K<n<10K

Danish Political Personas (FV26)

Multi-turn conversational persona dataset for 11 Danish political parties, generated using steering vectors on gpt-oss-20b.

Dataset Description

• Records: 2,013 conversations (183 per party × 11 parties)
• Turns: 3 turns per conversation (opening question + 2 journalist-generated follow-ups)
• Alignment: 88% (1,771/2,013 conversations fully aligned with party positions)
• Language: Danish
• Format: JSONL with messages field (system, user, assistant, user, assistant, user, assistant)

Parties

Question Categories

Pipeline

1. Data source: Kandidattest FV26 (29,050 candidate responses across 29 questions)
2. Contrastive pairs: 7.7M pairs generated from candidate answers
3. Steering vector training: gpt-oss-20b, layers 10-18, activation_mode="stream"
4. Layer optimization: Sweep confirmed layers 10-18 optimal
5. Multiplier optimization: Per-party sweep (0.5-0.9) → optimized multipliers
6. Multi-turn generation:
   • Varied system prompts (12 templates: TV debate, interview, podcast, town hall, etc.)
   • Opening question from expanded set
   • Follow-up questions generated by unsteered model (journalist persona)
   • Turn-1 alignment validated against party positions from candidate test
   • Misaligned responses regenerated with up to 2x multiplier
7. Output: 3-turn conversations in chat format

Record Format

{
  "messages": [
    {"role": "system", "content": "Du er spidskandidat for Socialdemokratiet..."},
    {"role": "user", "content": "Hvad er dit standpunkt til: ..."},
    {"role": "assistant", "content": "Som socialdemokrat mener jeg..."},
    {"role": "user", "content": "Men hvordan vil du konkret..."},
    {"role": "assistant", "content": "Vi vil..."},
    {"role": "user", "content": "Hvad hvis..."},
    {"role": "assistant", "content": "I det tilfælde..."}
  ],
  "party": "socialdemokratiet",
  "display_name": "Socialdemokratiet",
  "category": "political_statements",
  "question_id": 12,
  "turns": 3,
  "multiplier": 0.7,
  "aligned": true,
  "regenerated": false,
  "turn_details": [
    {"turn": 1, "aligned": true, "regenerated": false},
    {"turn": 2, "aligned": true, "regenerated": false},
    {"turn": 3, "aligned": true, "regenerated": false}
  ],
  "statement": "Det er vigtigere at kunne udvise flere kriminelle udlændinge end at overholde internationale konventioner",
  "expected_stance": "con"
}

Usage

from datasets import load_dataset

dataset = load_dataset("schneiderkamplab/danish-political-personas")

Files

• alternativet.jsonl — 183 conversations
• danmarksdemokraterne_inger_støjberg.jsonl — 183 conversations
• dansk_folkeparti.jsonl — 183 conversations
• det_konservative_folkeparti.jsonl — 183 conversations
• enhedslisten_de_rød-grønne.jsonl — 183 conversations
• liberal_alliance.jsonl — 183 conversations
• moderaterne.jsonl — 183 conversations
• radikale_venstre.jsonl — 183 conversations
• sf_-_socialistisk_folkeparti.jsonl — 183 conversations
• socialdemokratiet.jsonl — 183 conversations
• venstre_danmarks_liberale_parti.jsonl — 183 conversations

Citation

@misc{danish-political-personas-2026,
  title={Danish Political Personas: Multi-turn Conversational Dataset via Steering Vectors},
  author={JacobBITLABS},
  year={2026},
  publisher={schneiderkamplab},
  url={https://huggingface.co/datasets/schneiderkamplab/danish-political-personas}
}