danish-foundation-models/ai-arenaen-raw

AI-Arenaen: Danish-language conversations and human preferences

AI-Arenaen is a public chatbot arena for Danish users. People chat with two anonymous models side by side and say which answer they prefer. This dataset is the result.

Each row is one turn of a conversation: the two models' answers to the same user message, the preference the user gave on that turn (if any), and the full conversation both answers belong to. The prompts come from real users, most of them writing in Danish.

3,500+	~60%	90+	DA
paired responses	of turns rated by users	models compared	primary language

The data covers conversations since November 2025 and is updated as new conversations are collected.

How the data is collected

A user sends a prompt and gets back two answers from two anonymous models. The pair is usually drawn at random; the mode field records how it was chosen. The user can keep chatting and, at any point, react to a turn: model A was better, model B was better, both good, or both bad. The model names are revealed only once a preference is given.

Each conversation therefore leaves three things behind, and all three are in this dataset:

• the full exchange with each model (full_conversation_a / full_conversation_b),
• the user's preference per turn (choice),
• metadata: themes, languages, token counts, latency, and estimated energy use.

AI-Arenaen is the Danish version of compar:IA, the chatbot arena developed by the French Ministry of Culture. It is adapted and run for Danish users by Danish Foundation Models and the Danish Agency for Digital Government. Its aims are to raise awareness about model diversity, bias, and the environmental cost of conversational AI, and to release open Danish alignment data. The source code is available on GitHub.

Dataset structure

One default config, one train split, stored in ai-arenaen.parquet. A row is a single turn: the two models' answers to the same user message, plus the preference given on that turn. Rows from the same conversation share a comparison_id and run in turn order, so a conversation with three turns produces three rows. Most conversations are a single turn.

There is no held-out split. Note that full_conversation_a / full_conversation_b repeat across every row of a conversation, so split on comparison_id, not on rows, to avoid leaking the same conversation into both train and test.

Top-level columns

Column	Type	Description
response_id	string	Unique identifier for this turn-level entry (one row).
comparison_id	string	Identifier of the parent paired conversation. All turns of the same conversation share this id.
timestamp	datetime	When the conversation took place (e.g. 2025-11-13 09:09:54.563). Conversation-level: identical across all rows sharing a comparison_id.
turn	int	Zero-indexed position of this turn within the conversation.
model_a / model_b	string	Identifiers of the two compared models (anonymous to the user during the chat).
response_a / response_b	list of messages	The exchange for this turn only: the user message and each model's answer. Each message is an object with role, content, user_content (the user's message; null on assistant messages), and reasoning_content (the model's reasoning trace, when available).
full_conversation_a / full_conversation_b	list of messages	The complete conversation with model A / model B (all turns), same message schema as above. Identical across rows of the same comparison_id.
choice	string (nullable)	The user's preference for this turn. One of a_better, b_better, both_good, both_bad, idk, or null when no reaction was given.
metadata	struct	Per-turn and per-conversation metadata (see below).

metadata fields

Suffixes _a / _b refer to model A / model B. total_* fields describe the whole conversation; the un-prefixed counterparts describe this turn only.

Field	Type	Description
categories	list of string	Thematic categories of the conversation (e.g. Education, Politics & Government).
languages	list of string	Languages detected in the conversation (ISO 639-1 codes).
short_summary	string	Brief English summary of the conversation content.
mode	string	How the model pair was selected: random, custom (picked by the user), big-vs-small, or small-models.
custom_models_selection	list of string	Models picked by the user when they chose the pair manually.
tokens_a / tokens_b	int	Output tokens generated by each model for this turn.
total_tokens_a / total_tokens_b	int	Output tokens generated by each model over the whole conversation.
duration_a / duration_b	float	Generation duration (seconds) for this turn.
latency_a / latency_b	float	Latency (seconds, e.g. time to first token) for this turn.
time_to_vote	float (nullable)	Time in seconds the user took to react on this turn, from when the two answers were shown to when the preference was given. Only recorded since June 2026; null before that and when no reaction was given.
conso_a / conso_b	float	Estimated electricity consumption (kWh) for this turn.
total_conso_a / total_conso_b	float	Estimated electricity consumption (kWh) over the whole conversation.
participation_terms_version	string	Version of the terms of use the user accepted (legacy-pre-versioning for conversations collected before the terms were versioned).

Distribution of preferences (choice)

Value	Meaning	Share of turns
a_better	Model A preferred	~19%
b_better	Model B preferred	~19%
both_good	Both answers good	~15%
both_bad	Both answers bad	~5%
idk	Undecided	~2%
null	No reaction on this turn	~40%

About 60% of turns carry a preference. The rest have choice = null: these are unrated turns, useful as raw Danish conversation data even without a label.

Languages

About 9 in 10 turns are detected as Danish (da). The rest are mostly English, with some Norwegian, Swedish, French, German and others; a conversation can have more than one language.

Other files

• ai-arenaen_samples.jsonl / ai-arenaen_samples.tsv: a random sample of 1,000 rows in JSON Lines and TSV form, for a quick look without loading the parquet file.
• vote_tags.json: the reaction tags available on the platform (e.g. forkert_sprog / "Incorrect language", faktuelt_forkert / "Factually incorrect"), with their sign and Danish and English labels.

Quick start

from datasets import load_dataset

ds = load_dataset("danish-foundation-models/ai-arenaen", split="train")

# Build (prompt, chosen, rejected) pairs from the clearly-preferred turns
def to_pair(row):
    chosen, rejected = ("response_a", "response_b") if row["choice"] == "a_better" else ("response_b", "response_a")
    return {"chosen": row[chosen], "rejected": row[rejected]}

pairs = (ds.filter(lambda r: r["choice"] in {"a_better", "b_better"})
           .map(to_pair))

# Or pull back a whole conversation from any of its rows
convo = ds.filter(lambda r: r["comparison_id"] == ds[0]["comparison_id"])

What it is good for

The obvious use is preference tuning (DPO, reward models) and evaluation of Danish chat models. Beyond that, the unrated conversations are a sample of how Danish speakers actually use chatbots, and the per-turn energy and token figures let you study cost and efficiency across 90+ models on the same prompts.

If you build something with it, we would like to hear about it: open a discussion on the dataset discussion page.

Privacy and content

Users consent through the site's terms of use. We ran PII detection and excluded conversations that contained personal information. We do not filter toxic or hateful content: it stays in so that safety and moderation can be studied on real data, so expect some.

Licenses

The dataset is released under the open Etalab 2.0 and CC-BY-4.0 licenses, subject to potential third-party claims regarding language-model outputs. It is the responsibility of users to ensure their use complies with applicable laws and regulations, in particular regarding data protection and the terms of use of the different model providers.

Citation

If you use this dataset, please cite:

@misc{aiarenaen,
  title        = {AI-Arenaen: A Danish Chatbot Arena Dataset},
  author       = {{Danish Foundation Models}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/danish-foundation-models/ai-arenaen}},
}

Reporting sensitive data

If you find a row that you believe contains PII or sensitive data, please let us know via this short form.

Contact

For any question or request, open a discussion on the dataset discussion page.

---