SeaSlidesBench-Rich

SeaSlidesBench-Rich is a 32-task stress-test extension of SeaSlidesBench for evaluating slide generation systems on long slide decks, mathematical formulas, code, pseudocode, and visualizations. It contains 27 English tasks, 5 Chinese tasks, and 10 arXiv paper attachments. All tasks specify a 16:9 aspect ratio.

This dataset provides task inputs only; it does not include reference slides, model outputs, or evaluation scores. This release preserves the original task text, fields, tags, and ordering to support reproduction of existing experiments.

Contents

• Quick Start
• Files and Task IDs
• Fields
• Dataset Statistics
• Paper Sources
• Known Limitations
• License

Quick Start

Install the dependencies: pip install datasets huggingface_hub.

from pathlib import Path

from datasets import load_dataset
from huggingface_hub import snapshot_download

repo_id = "OrangeX4/SeaSlidesBench-Rich"
# For reproducible experiments, replace main with a fixed dataset commit hash.
revision = "main"
dataset = load_dataset(repo_id, split="validation", revision=revision)

# load_dataset loads the task table; snapshot_download also downloads the PDFs.
root = Path(snapshot_download(repo_id, repo_type="dataset", revision=revision))
for index, sample in enumerate(dataset):
    task_id = 129 + index
    attachments = [root / path for path in sample["attachments"]]
    assert all(path.is_file() for path in attachments)
    print(task_id, sample["prompt"], attachments)

If you have already downloaded the entire directory, you can read the JSONL file directly:

import json
from pathlib import Path

root = Path("SeaSlidesBench-Rich")  # Set this to your local download directory.
with (root / "data/valid.jsonl").open(encoding="utf-8") as stream:
    samples = [json.loads(line) for line in stream if line.strip()]

Files and Task IDs

SeaSlidesBench-Rich/
├── README.md
├── LICENSE
├── manifest.json          # File hashes, sizes, and attachment source links
└── data/
    ├── valid.jsonl        # 32 tasks, one JSON object per line
    └── arxiv/             # 10 paper PDFs

The dataset has a single validation split and no training or test splits. This name follows the existing file naming convention; it does not imply an independently established train/validation/test partition.

JSONL rows 1–32 correspond to task IDs 129–160 in the combined SeaSlidesBench evaluation: task_id = row_number (1-based) + 128. The records do not have a separate ID field, so preserve their original order. The first 128 tasks are not included here.

All attachments paths are relative to the dataset root, not to the data/ directory. File checksums are provided in manifest.json.

Fields

Field	Type	Description
prompt	string	Original slide generation instruction
language	string	Output language: en or zh
source	string	Task category: lecture, arXiv, or code_repo; see the table below
aspect_ratio	string	Slide aspect ratio; 16:9 for all tasks
page_low	int	Inclusive lower bound of the original evaluation slide-count range
page_high	int	Inclusive upper bound of the original evaluation slide-count range
attachments	list[string]	Relative attachment paths; an empty list when there are no attachments
extra_info.arxiv_id	string or null	Paper's arXiv ID; null for non-paper tasks
extra_info.tags	list[string]	Original capability tags; a task can have multiple tags
extra_info.scenario	string	Intended presentation scenario
extra_info.audience	string	Intended audience

The structured slide-count ranges do not always match the natural-language requirements in the prompts. See Known Limitations for the specific discrepancies.

Dataset Statistics

Task category	Count	Inputs provided
lecture	19	Course or teaching-topic prompts; no source lecture notes are attached
arXiv	10	Paper presentation prompts with the corresponding PDFs
code_repo	3	Technical-topic prompts about Rust, PyTorch, and RESTful APIs; no repository attachments, URLs, or pinned commits are provided

source is a legacy task category, not a complete provenance record. In particular, lecture and code_repo do not mean that actual lecture notes or code repositories are included.

Capability tag	Tasks	Percentage
long_document (long slide decks)	15	46.9%
math (mathematical content)	23	71.9%
code (code blocks)	13	40.6%
pseudocode	11	34.4%
visualization	25	78.1%

Tags overlap: 20 tasks have at least one of the code and pseudocode tags, and each task has 2.72 tags on average. In this dataset, long_document identifies the 15 tasks with page_low >= 40; it describes the target slide-deck length, not the length of an input PDF. The original tags are preserved and may not exhaustively cover every capability mentioned in a prompt.

Paper Sources

arXiv ID	Paper
1301.3781	Efficient Estimation of Word Representations in Vector Space (Word2Vec)
1312.6114	Auto-Encoding Variational Bayes (VAE)
1406.2661	Generative Adversarial Nets (GAN)
1609.02907	Semi-Supervised Classification with Graph Convolutional Networks (GCN)
1706.03762	Attention Is All You Need (Transformer)
1707.06347	Proximal Policy Optimization Algorithms (PPO)
2002.05709	A Simple Framework for Contrastive Learning of Visual Representations (SimCLR)
2010.11929	An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (ViT)
2106.09685	LoRA: Low-Rank Adaptation of Large Language Models
2305.18290	Direct Preference Optimization: Your Language Model is Secretly a Reward Model (DPO)

The original records do not include arXiv version numbers or download dates. Later versions on arXiv may differ from the included attachments. For reproducibility, pin the dataset commit and verify attachment hashes against manifest.json.

Known Limitations

• This is a small, English-heavy set of 32 stress-test tasks for rich-content generation. It is not representative of all real-world presentation needs.
• This release does not include the original candidate pool, selection logs, or independent annotation-agreement records. It should not be taken as evidence of unbiased sampling or independent human validation.
• The 22 tasks without attachments require systems to organize content from a topic prompt. Their information-access conditions differ from those of the paper-based tasks. When comparing systems, report whether retrieval is allowed and which external materials are used.
• Some natural-language minimum slide counts are stricter than page_low. The original values are retained to avoid changing existing experimental conditions. When reproducing results, report compliance with the prompt requirements separately from compliance with the structured slide-count ranges.

JSONL row	Combined evaluation task ID	Prompt requirement	Original slide-count range
3	131	At least 55 slides	50–60
5	133	At least 45 slides	42–50
7	135	At least 50 slides	48–55
13	141	At least 50 slides	48–55

License

The dataset's own tasks, metadata, and documentation are released under the MIT License.

Third-party papers in data/arxiv/ are not covered by this repository's MIT license. They remain subject to the copyrights and applicable licenses of their original authors or rights holders. Consult the source links above for each paper's license information; this dataset's MIT label does not grant permission to relicense or freely redistribute those papers. See the official arXiv licensing guidance.