swiss-ai/Apertus-v1.5-Preference-Pool

Apertus 1.5 Preference Pool

The following dataset contains responses generated from a pool of 24 models, alongside the judge annotations that score them according to rubric-based rewards. The dataset was used for constructing the final Apertus 1.5 preference dataset, Apertus-v1.5-Preference-Data.

It is derived from Ai2's Olmo 3 Dolci-Instruct-DPO, where we reuse only the prompts and regenerate every response.

Where the preference dataset ships a single chosen / rejected pair per prompt, this dataset ships all 24 model responses per prompt together with their judge annotations, so the pair-construction step can be redone — max-min, max-random, or any other acquisition rule — without regenerating or re-annotating anything.

The pipeline behind it, response generation and annotation alike, follows ActiveUltraFeedback (code).

How this dataset was built

• Prompts. Taken from Dolci-Instruct-DPO (ODC-BY).
• Response generation. One response per prompt from each of the 24 models listed at the end of this card, generated with the swiss-ai/posttraining-data repository.
• Annotation. Every response was scored by Qwen3-235B-A22B-Instruct-2507 as judge. The judge rates four aspects — instruction_following, honesty, truthfulness, helpfulness — and each aspect is kept as a probability distribution over the 1-5 scale. The judge model is also one of the 24 response models.

Dataset structure

One row per prompt:

field	type	description
prompt_id	string	Dolci-Instruct-DPO prompt id
prompt	list[{role, content}]	the conversation up to and including the final user turn
completions	list[{model, response, annotation}]	24 entries, one per model

Every row carries all 24 responses, and they appear in the same model order in every row, so index i refers to the same model throughout the dataset.

Each annotation holds four aspects, each mapping the scores "1"-"5" to probabilities that sum to 1:

{
  "instruction_following": {"1": 7.5e-06, "2": 2.3e-05, "3": 0.00098, "4": 0.0474, "5": 0.9516},
  "honesty":               {...},
  "truthfulness":          {...},
  "helpfulness":           {...},
}

A scalar score is the weighted sum sum(int(k) * p for k, p in aspect.items()).

Size. 257,953 prompts, each carrying all 24 responses — 6,190,872 responses in total.

Usage

from datasets import load_dataset

ds = load_dataset("swiss-ai/Apertus-1.5-Preference-Pool", split="train")

row = ds[0]
for c in row["completions"]:
    helpfulness = sum(int(k) * p for k, p in c["annotation"]["helpfulness"].items())
    print(c["model"], round(helpfulness, 3))

Some of what this is useful for:

• building preference pairs under any acquisition rule;
• training reward models or evaluators on the pooled responses and their aspect scores;
• studying judge behaviour — score calibration, how the aspects correlate, and how the distributions shift across models of very different capability.

License & attribution

This dataset is a derivative of Dolci-Instruct-DPO and is released under the ODC-BY license, in accordance with the source dataset, following Ai2's Responsible Use Guidelines.

The responses were generated by the 24 models listed below, each under its own license. All are permissive: Apache-2.0 throughout, except Phi-4-mini-instruct (MIT) and Trinity-Nano-Preview (OpenMDW-1.1). Per-model licenses and links are in the table below.

Responses are model outputs and carry no human review; annotations are produced by an LLM judge.

Citation

If you use this dataset, please cite the Apertus project:

@misc{apertus2025,
  title  = {TODO: Apertus citation},
  author = {TODO},
  year   = {TODO},
  url    = {TODO},
}

The pipeline behind this dataset is described in:

@misc{activeultrafeedback2026,
  title  = {ActiveUltraFeedback: Efficient Preference Data Generation using Active Learning},
  author = {Melikidze, Davit and Schneider, Marian and Lam, Jessica and Wertich, Martin
            and Hakimi, Ido and P{\'a}sztor, Barna and Krause, Andreas},
  year   = {2026},
  eprint = {2603.09692},
  archivePrefix = {arXiv},
  url    = {https://arxiv.org/abs/2603.09692},
}

Model pool

The 24 models whose responses make up this dataset, with the license each is released under.

Model	Hugging Face repo	License
EuroLLM-1.7B-Instruct	utter-project/EuroLLM-1.7B-Instruct	apache-2.0
EuroLLM-9B-Instruct-2512	utter-project/EuroLLM-9B-Instruct-2512	apache-2.0
EuroLLM-22B-Instruct-2512	utter-project/EuroLLM-22B-Instruct-2512	apache-2.0
Ministral-3-3B-Instruct-2512	mistralai/Ministral-3-3B-Instruct-2512	apache-2.0
Ministral-3-8B-Instruct-2512	mistralai/Ministral-3-8B-Instruct-2512	apache-2.0
Ministral-3-14B-Instruct-2512	mistralai/Ministral-3-14B-Instruct-2512	apache-2.0
Mistral-Small-24B-Instruct-2501	mistralai/Mistral-Small-24B-Instruct-2501	apache-2.0
Mistral-Large-3-675B-Instruct-2512	mistralai/Mistral-Large-3-675B-Instruct-2512	apache-2.0
Mixtral-8x22B-Instruct-v0.1	mistralai/Mixtral-8x22B-Instruct-v0.1	apache-2.0
Phi-4-mini-instruct	microsoft/Phi-4-mini-instruct	mit
Qwen2.5-0.5B-Instruct	Qwen/Qwen2.5-0.5B-Instruct	apache-2.0
Qwen2.5-1.5B-Instruct	Qwen/Qwen2.5-1.5B-Instruct	apache-2.0
Qwen3-0.6B	Qwen/Qwen3-0.6B	apache-2.0
Qwen3-1.7B	Qwen/Qwen3-1.7B	apache-2.0
Qwen3-4B-Instruct-2507	Qwen/Qwen3-4B-Instruct-2507	apache-2.0
Qwen3-8B	Qwen/Qwen3-8B	apache-2.0
Qwen3-32B	Qwen/Qwen3-32B	apache-2.0
Qwen3-30B-A3B-Instruct-2507	Qwen/Qwen3-30B-A3B-Instruct-2507	apache-2.0
Qwen3-Next-80B-A3B-Instruct	Qwen/Qwen3-Next-80B-A3B-Instruct	apache-2.0
Qwen3-Omni-30B-A3B-Instruct	Qwen/Qwen3-Omni-30B-A3B-Instruct	apache-2.0
Qwen3-235B-A22B-Instruct-2507	Qwen/Qwen3-235B-A22B-Instruct-2507	apache-2.0
Qwen3.5-397B-A17B	Qwen/Qwen3.5-397B-A17B	apache-2.0
SmolLM3-3B	HuggingFaceTB/SmolLM3-3B	apache-2.0
Trinity-Nano-Preview	arcee-ai/Trinity-Nano-Preview	OpenMDW-1.1