Datasets:
license: cc-by-nc-sa-4.0
task_categories:
- visual-question-answering
- image-text-to-text
language:
- en
- zh
size_categories:
- 10K<n<100K
tags:
- preview
- multimodal
- scientific-reasoning
- physics
- chemistry
- biology
- mathematics
configs:
- config_name: default
default: true
data_files:
- split: train
path: data/train-*.parquet
dataset_info:
features:
- name: id
dtype: string
- name: images
list: image
- name: problem
dtype: string
- name: matched_solution
dtype: string
splits:
- name: train
num_examples: 41828
Dataset Card for SCP-VL — Preview
Preview release (v0.1-preview, September 2026). This dataset is an early release and has known quality issues. Problems, images, and solutions are automatically extracted and matched. Information may be missing or corrupted; attached images may be incorrect, incomplete, or unrelated; solutions may contain errors or omit required answers. Further corrections and quality improvements are planned. Do not treat this release as a verified benchmark or its solutions as guaranteed ground truth.
当前为预览版。 数据由自动抽取与匹配流程生成，可能存在题目信息缺失、裁图不完整、错配或多配图片、公式/化学结构丢失、题解错误等问题。后续将逐步修订；使用前请按具体任务进行检查。
Dataset Description
Paper
Dataset Summary
SCP-VL is a multimodal extension of the SCP scientific problem-solution dataset family. This preview contains 41,828 problem-solution pairs, each with at least one attached image and a text solution. Images include scientific diagrams, plots, tables, and other visual material associated with the problems.
The dataset spans physics, chemistry, biology, mathematics, and related subjects. It contains English and Chinese material, with educational levels ranging from secondary-school and competition problems to university-level content. An attached image does not necessarily mean that the problem requires visual reasoning to solve.
This release contains 51,405 image attachments across the 41,828 examples; some images occur in multiple examples. Images are embedded in the Parquet files and can be loaded directly with Hugging Face Datasets. Separate image downloads and local source paths are not required.
GitHub: SCP dataset pipeline
Related text-only dataset: EricLu/SCP-378K
Supported Tasks
Potential research uses include:
- Scientific visual question answering and multimodal reasoning.
- VLM instruction tuning after task-specific validation and filtering.
- Reward-model or verifier research using the extracted reference solutions, subject to their known errors.
Solutions may be short answers rather than complete derivations. Suitability for supervised training or evaluation should be assessed separately for the intended use.
Languages
English and Chinese.
Dataset Structure
Data Fields
The public dataset contains exactly four columns:
| Field | Type | Description |
|---|---|---|
id |
string | Opaque example identifier, such as scpvl_ followed by 32 hexadecimal characters. It does not encode a source title, source path, or page number. |
images |
list of Image |
One or more attached images, embedded as bytes in the Parquet files. Hugging Face Datasets decodes these into a list of PIL images by default. |
problem |
string | Extracted problem text, which may contain mathematical notation and references to figures or tables. |
matched_solution |
string | Automatically matched text solution. It may contain OCR errors, incomplete reasoning, or incorrect results. |
Source-title fields, source identifiers, source filenames, and page metadata are not included. Embedded images have no source paths or filenames in the image feature. Text link destinations and explicit occurrences of the corresponding source-title string have been removed during export. Scientific content and printed text within the crops are otherwise retained; this is not a claim that source material cannot be recognized from its content.
Data Splits
| Split | Examples |
|---|---|
train |
41,828 |
This is a single release split. No separate validation or test split, or contamination-free evaluation partition, is provided.
Loading the Dataset
from datasets import load_dataset
dataset = load_dataset(
"EricLu/SCP-VL",
split="train",
revision="v0.1-preview",
)
example = dataset[0]
print(example["id"])
print(example["problem"])
print(example["matched_solution"])
print(len(example["images"]))
image = example["images"][0] # PIL image
For streaming access:
dataset = load_dataset(
"EricLu/SCP-VL",
split="train",
revision="v0.1-preview",
streaming=True,
)
example = next(iter(dataset))
Relationship to SCP-378K
SCP-378K contains 377,705 text problem-solution pairs. SCP-VL publishes the image-bearing subset recovered through the newer multimodal pipeline as a separate dataset.
The multimodal pipeline reprocessed problems with visual references before the earlier text-only filtering stages. It recovered 46,212 usable-by-model examples, of which 41,828 have attached images and are included here. The remaining 4,384 examples without images are excluded from this release.
SCP-VL is not a disjoint set of new questions relative to SCP-378K: some questions overlap or replace earlier text-only versions. Do not simply concatenate the two datasets without checking overlap. The 417,149-example merged dataset is not included in this repository. SCP-378K's published data remain unchanged.
Dataset Creation
Source Data
The dataset was produced from scientific and educational documents using an automated pipeline. Public records contain the problem, associated visual material, and a matched solution, without source-title metadata.
Extraction and Matching
The pipeline includes:
- Text recognition, segmentation, and extraction of problems and candidate solutions.
- Figure/table detection and image cropping.
- Extraction of visual references from each problem.
- Retrieval of candidate figures within the corresponding document, followed by model-based image matching.
- Solution matching with the selected images supplied to a vision-language model.
- Model-based usability filtering.
This preview uses WeMM-Embedding-9B for figure retrieval and Qwen3.8-27B for the main visual matching and filtering stages. More than one image can be retained for a problem. These automated checks do not guarantee that all retained images or solutions are correct.
Preview Quality Check
A fixed random sample of 100 image-bearing examples, containing 117 images, was reviewed by a Codex assistant through direct inspection of the problem text, images, and solution text. Selected cases were compared with the original document pages.
| Review outcome | Examples |
|---|---|
| Clear issues requiring correction or exclusion | 30 |
| Main answer remained usable in the review, but revisions were needed | 26 |
| No obvious issue found during this review | 44 |
This was an AI-assisted spot check, not independent human expert annotation. Complex derivations were not exhaustively verified. These counts describe the sample; they are not a precise error-rate estimate for the complete dataset. “No obvious issue found” does not mean an example is fully correct.
The identified problems have not been comprehensively corrected in this preview. Apart from removing source-related export metadata and link destinations, this release retains the current extracted problem/image/solution content, including known problematic examples.
Considerations for Using the Data
Known Limitations
- Missing problem information: some examples lack an explicit question, shared instructions, or necessary context from another problem.
- Image matching errors: a problem may include a wrong image or extra images belonging to other problems.
- Incomplete crops: scale bars, numerical labels, or other essential details may be cut off.
- Text extraction errors: mathematical expressions, chemical structures, tables, and symbols may be lost, corrupted, or inconsistently represented in text and images.
- Solution errors: answers may be incomplete, internally inconsistent, or wrong, including mistakes already present in the source material.
- Answer information in images: some crops include worked-example annotations or solution diagrams, which can reveal part of the answer.
- Formatting noise: residual placeholders, literal newline escapes, repeated content, or redundant visual descriptions may remain.
- Uneven coverage: subjects, languages, difficulty levels, and source quality are not uniformly distributed.
- Overlap: overlap with related SCP releases or external training/evaluation datasets has not been eliminated for this release.
Feedback and Updates
Please report issues through the dataset's Community tab, including the example id, the observed problem, and any proposed correction. The opaque ID is intended to support tracking corrections across releases. Future releases may correct or remove examples; IDs of retained examples will be preserved where possible.
Use the v0.1-preview tag or a commit revision for reproducible experiments. The main branch may receive subsequent corrections.
Additional Information
Dataset Curators
The dataset was created as part of research on scientific reasoning and automated problem-solution extraction.
Licensing Information
This dataset is released under the CC-BY-NC-SA 4.0 license, consistent with the SCP-378K release.
Citation Information
If you use this dataset, please cite the SCP pipeline paper and specify the SCP-VL version used:
@misc{lu2025scp116khighqualityproblemsolutiondataset,
title={SCP-116K: A High-Quality Problem-Solution Dataset and a Generalized Pipeline for Automated Extraction in the Higher Education Science Domain},
author={Dakuan Lu and Xiaoyu Tan and Rui Xu and Tianchu Yao and Chao Qu and Wei Chu and Yinghui Xu and Yuan Qi},
year={2025},
eprint={2501.15587},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2501.15587},
}