ActiveUltraFeedback — Combined

This is a preference dataset of 140k samples generated for the paper ActiveUltraFeedback: Efficient Preference Data Generation using Active Learning (Melikidze et al., 2026).

The prompts are from a combination of the version of UltraFeedback (Cui et al., 2023) that was released by AllenAI (at allenai/ultrafeedback_binarized_cleaned) and Skywork Reward Preference 80k v0.2 (Liu et al., 2024). The response pairs were generated with the ActiveUltraFeedback pipeline, which calls a large pool of open-weight LLMs to first generate candidate responses, then uses various active selection strategies to choose which response pair per prompt to select for preference annotation by an LLM-as-a-Judge.

Configs

Each of the 11 configs is a train split. The largest is called full, and it contains all 30 responses that we generated for each prompt. The remaining 10 splits have only two responses per prompt, and the config names are of this form: <preference_optimization_method>_<response_pair_selection_method>.

Preference optimization method (2 options):

• rm — Reward Modeling
• dpo — Direct Preference Optimisation

Response pair selection method (5 options): The first three are based on existing works, and the final two are novel.

• infomax — InfoMax (Saha, 2021) prioritizes pure exploration by selecting the response pair with the highest joint uncertainty, regardless of the predicted reward quality.
• dts — DTS (Double Thompson Sampling) (Wu & Lin, 2016) addresses the exploration-exploitation trade-off by drawing two independent samples from the reward posterior and selecting the responses that maximize them.
• maxminlcb — MaxMinLCB (Pásztor et al., 2024) considers pairwise Lower Confidence Bounds (LCBs) and selects the response pair where the first response maximizes the minimum LCB against all responses, and the second response minimizes the LCB against the first response.
• deltaucb — DeltaUCB identifies pairs with the largest optimistic quality difference, i.e., the pair that maximizes the probability that the first response is preferred over the second response in the best-case scenario.
• drts — DRTS (Double Reversed Thompson Sampling) selects one response that maximizes and another that minimizes their respective samples from the reward posterior. This strategy explicitly targets pairs with a significant delta in quality, while the underlying stochastic sampling preserves exploration and diversity.

Features

The features for the full config are listed below:

Field	Type	Description
prompt_id	string	Unique prompt identifier
prompt	string	Input prompt
prompt_licenses	list of strings	Licenses applicable to the input prompt
source	string	Where the original dataset obtained the prompt from.
completions	list of dict	Each dict contains a model name, the response generated by the model (along with the system prompt and messages for generation), and how good the response is on a given principle according to an LLM Judge.

The remaining 10 configs have these features:

Field	Type	Description
prompt_id	string	Unique prompt identifier
prompt	string	Input prompt
prompt_licenses	list of strings	Licenses applicable to the input prompt
chosen	list of {role, content}	Preferred response
chosen_model	string	Model that generated the chosen response
chosen_score	float	Reward model score for the chosen response
chosen_licenses	list of strings	Licenses applicable to the chosen response
rejected	list of {role, content}	Dispreferred response
rejected_model	string	Model that generated the rejected response
rejected_score	float	Reward model score for the rejected response
rejected_licenses	list of strings	Licenses applicable to the rejected response

Licenses

The prompts and responses are derived from third-party sources under their respective licenses. We have annotated every prompt and response with the licenses that may apply. Review all constituent licenses before use. See LICENSES.md for the full list.

Citation

If you use this dataset, please cite:

@misc{melikidze2026activeultrafeedbackefficientpreferencedata,
      title={ActiveUltraFeedback: Efficient Preference Data Generation using Active Learning}, 
      author={Davit Melikidze and Marian Schneider and Jessica Lam and Martin Wertich and Ido Hakimi and Barna Pásztor and Andreas Krause},
      year={2026},
      eprint={2603.09692},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2603.09692}, 
}