UltraPresent

Project: https://github.com/icip-cas/PPTAgent

Paper: https://arxiv.org/abs/2602.22839

UltraPresent is a dataset for automatic presentation generation, designed to support research on large language model (LLM) and agent-based systems that generate structured slide presentations from natural language instructions.

The dataset contains a collection of diverse presentation generation tasks, where each example specifies a user instruction describing the topic, structure, and constraints of a presentation. These tasks aim to evaluate and train systems capable of producing coherent, structured slide decks that combine content organization, narrative planning, and presentation design.

Presentation generation is a complex multimodal task that requires not only content understanding and summarization but also the ability to structure information into slides and maintain logical coherence across the entire deck. Recent work such as DeepPresenter frames presentation generation as a long-horizon agentic process involving information research, slide drafting, and iterative refinement based on rendered artifacts.  UltraPresent is designed to support such research by providing diverse, realistic prompts that simulate real-world presentation requests.

Dataset Characteristics

•	Task type: Instruction-based presentation generation
•	Scale: ~10K samples
•	Languages: English and Chinese

Each instance typically includes:

•	Instruction/prompt: A natural-language description of the desired presentation.
•	Constraints: Requirements such as slide count, aspect ratio, structure, or target audience.

Loading the Dataset:

from datasets import load_dataset

dataset = load_dataset("Forceless/UltraPresent")