HPDv3 Human Preference Dataset

## Overview Human Preference Dataset v3 (HPDv3) is a large-scale dataset for training and evaluating human preference models over text-to-image generations. It contains 1.08M text-image pairs and 1.17M annotated pairwise preference comparisons, distributed in JSON format. Released alongside the HPSv3 paper (ICCV 2025) by Mizzen AI and CUHK MMLab, it targets wide-spectrum preference modeling across image quality, prompt alignment, and aesthetics. ## Schema - prompt — string — text prompt used to generate the image - image — string/path — reference to the generated image - model — string — generator model that produced the image - pair_id — string — identifier linking the two images in a pairwise comparison - preferred — int — which image in the pair was preferred by annotators - annotator_id — string — anonymized annotator identifier - score — float — preference score / rating (where available) - +additional metadata columns for source, split, and annotation provenance ## Sources - HuggingFace: https://huggingface.co/datasets/MizzenAI/HPDv3 — MIT license - Paper: arXiv:2508.03789 (HPSv3: Towards Wide-Spectrum Human Preference Score, ICCV 2025) ## Methodology The dataset aggregates text-image pairs from multiple text-to-image generators across a wide spectrum of prompt domains and quality levels. Pairwise preference annotations were collected from human annotators comparing two generations for the same or related prompts, with the goal of training reward/preference models (HPSv3) that generalize across styles, qualities, and prompt types. The authors emphasize 'wide-spectrum' coverage to reduce bias toward any single generator family. ## Known gaps & limitations Language coverage is English-only. Human preference data inherits annotator demographic and cultural biases that the source does not fully document. Image generator coverage is a snapshot in time and will become stale as new models are released. Source does not document detailed annotator agreement statistics; buyers training reward models should validate calibration empirically and check for leakage against common T2I eval benchmarks. ## Intended use & out-of-scope - IS for: training human preference / reward models for text-to-image generation, RLHF/DPO fine-tuning of diffusion models, building T2I evaluation benchmarks, and research on preference modeling. - NOT for: deployment as a sole arbiter of image quality without bias auditing; not deduplicated against external T2I eval suites — leakage risk if used for benchmark training. _Federated dataset: 2 parquet shards, 248.4 MB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: HPDv3 — Human Preference Dataset v3 (1.08M text-image pairs) Wide-spectrum human preference dataset for text-to-image generation: 1.08M text-image pairs and 1.17M pairwise human preference annotations. MIT-licensed, used to train HPSv3 (ICCV 2025) reward models.