README.md

metadata

pretty_name: ViPPS
language:
  - vi
task_categories:
  - visual-question-answering
  - question-answering
tags:
  - physics
  - education
  - multimodal
  - vision-language
  - reasoning
  - chain-of-thought
  - vietnamese
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet

ViPPS: Vietnamese Physics Problem Solving

ViPPS is the first multimodal dataset for physics problem solving in Vietnamese. It contains 5,520 secondary- and high-school physics problems, each with a final answer, a unit, and a step-by-step chain-of-thought (CoT) solution. 3,605 problems are accompanied by an image (circuit diagrams, graphs, experimental setups, etc.); the remaining 1,915 are text-only.

• Paper: ViPPS: Building a Multimodal Dataset for Physics Problem Solving in Vietnamese (MIWAI 2025, Springer LNCS 16354)
• Language: Vietnamese (vi)
• Domain: Electricity, magnetism, AC circuits, electromagnetic induction, oscillations

Abstract

Recent advances in multimodal large models, particularly Vision-Language Models (VLMs), have demonstrated strong capabilities in handling reasoning tasks that involve both textual and visual information. However, in low-resource languages such as Vietnamese, research in educational contexts remains limited due to the absence of suitable multimodal datasets. To address this gap, we introduce ViPPS (Vietnamese Physics Problem Solving), the first multimodal dataset designed for physics problem solving in Vietnamese. ViPPS consists of nearly 500 physics-related images (e.g., circuit diagrams, mechanics illustrations, experimental setups, and graphs) paired with more than 5,500 corresponding textual questions collected from Vietnamese educational Q&A platforms. Each entry reflects authentic student learning scenarios in secondary and high school physics. We detail the dataset creation pipeline, cleaning procedures, and provide comprehensive statistical analyses of linguistic and visual characteristics. To benchmark ViPPS, we fine-tune recent VLM baselines under a supervised multimodal setting, and report their performance across different physics domains. Results highlight both the potential of current VLMs and their limitations in solving domain-specific multimodal reasoning problems in Vietnamese.

Repository Structure

vipps/
├── data/
│   └── train-00000-of-00001.parquet   # all 5,520 problems, images embedded
├── vipps.csv      # same records, image column = relative path (UTF-8 with BOM)
├── vipps.json     # same records as a JSON array
└── images/        # 3,605 raw image files referenced by `vipps.csv` / `vipps.json`

The Parquet file is the primary format: the image column is a decoded Image feature (shown directly in the Data Studio viewer). vipps.csv and vipps.json contain the same records, with the image column holding paths relative to the repository root (e.g. images/CH111.png).

Data Fields

Field	Type	Description
id	string	Unique problem identifier: category prefix + number (e.g. CH001, DLO256).
category	string	Problem category, one of 14 codes (see Categories).
modality	string	text_only or multimodal.
question	string	Problem statement in Vietnamese.
image	Image | null	Accompanying image (embedded in Parquet; a relative path such as images/CH111.png in the CSV/JSON). null for text_only problems.
answer	string	Final answer. Numeric for most problems; a few are expressions (e.g. 100 * pi) or categorical (Có / Không, in CHLT).
unit	string | null	Unit of the answer (e.g. A, V, Ω, W, V/m, %, N); empty for dimensionless or non-numeric answers.
cot	string	Step-by-step solution, typically structured as: data identification → formulas → substitution and calculation → conclusion.

Example

{
  "id": "CH001",
  "category": "CH",
  "modality": "text_only",
  "question": "Mạch RLC cộng hưởng, tổng trở đo được Z=40 Ω, hãy xác định điện trở thuần R?",
  "image": null,
  "answer": "40",
  "unit": "Ω",
  "cot": "1. Nhận diện dữ kiện:\n- Mạch RLC cộng hưởng, tổng trở Z = 40 Ω.\n\n2. Công thức:\n- Khi mạch RLC cộng hưởng, tổng trở đạt giá trị nhỏ nhất và bằng điện trở thuần R.\n- Tức là Z = R.\n\n3. Thay số và tính toán:\n- Z = R = 40 Ω.\n\n4. Kết luận:\n- Điện trở thuần R = 40 Ω."
}

Categories

Category	Description	Total	Text-only	Multimodal	Example ID
THCB	Experimental tasks: error analysis in circuits	400	80	320	THCB012
DT	Electric field vectors	400	69	331	DT045
LD	Superposition of electric forces/fields	400	400	0	LD078
TD	Parallel-plate capacitors, geometry-based	469	177	292	TD103
DLO	Circuits analyzed with Ohm's law (with diagrams)	808	0	808	DLO256
RLC	Series RLC circuits, compute U, I, P	484	0	484	RLC023
VI	Voltage–current oscillation graphs	400	0	400	VI145
NL	Electric and magnetic field energy (with graphs)	400	190	210	NL032
CH	Resonance: schematic + Z, f, R, L, C values	380	290	90	CH087
CHLT	Theoretical resonance analysis	20	20	0	CHLT009
DDT	LC oscillations and solenoid circuits	400	130	270	DDT041
CUT	Electromagnetic induction	261	0	261	CUT075
CUTLT	Theoretical electromagnetic induction	139	0	139	CUTLT008
QA	Posts from Q&A sites	559	559	0	QA400
Total		5,520	1,915	3,605

Dataset Statistics

Statistic	Value
Problems	5,520
Multimodal / text-only	3,605 / 1,915
Image files	3,605 (PNG 3,203 · JPG 365 · GIF 37)
Distinct image contents	337 (see note below)
Question length (words)	mean 46.1 · median 34 · max 240
CoT length (words)	mean 121.5 · median 101 · max 499
Numeric answers	4,618
Problems without a unit	427

Note on images. Many problems share the same figure with different numerical values, so the 3,605 image files contain only 337 distinct images (verified by file hash). This differs from the "nearly 500 images" figure in the paper abstract, which may count images at a different stage of the pipeline. When building train/validation/test splits, split by image content rather than by problem to avoid leakage between splits.

Usage

Load with 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("quin210/vipps", split="train")
print(ds[0])

ds[0]["image"] is a PIL.Image (or None for text-only problems):

ex = next(x for x in ds if x["image"] is not None)
ex["image"].show()

Load with pandas

import pandas as pd

df = pd.read_csv("vipps.csv", encoding="utf-8-sig")  # image column = file path
df.groupby(["category", "modality"]).size()

Filter by modality

text_only  = ds.filter(lambda x: x["modality"] == "text_only")
multimodal = ds.filter(lambda x: x["modality"] == "multimodal")

Intended Uses

• Training and evaluating Vision-Language Models on physics reasoning in Vietnamese.
• Supervised fine-tuning with chain-of-thought supervision (cot) and final-answer evaluation (answer + unit).
• Research on multimodal reasoning, educational AI, and low-resource languages.

Limitations

• Domain scope. The dataset focuses on electricity, magnetism, and related topics; it does not cover all of secondary- and high-school physics.
• Answer format. answer is stored as a string. Most are plain numbers, but some are expressions or Có/Không; evaluation scripts should normalize numerals (including decimal commas) and handle these cases.
• Image redundancy. Several problems reuse the same figure (see the note above).
• Splits. The release is provided as a single table without official train/validation/test splits.

Citation

If you use ViPPS, please cite:

@InProceedings{10.1007/978-981-95-4960-3_25,
author="Vo, Quynh T. N.
and Le, Xinh T.
and Tran, Thao H. M.
and Quan, Tho T.",
editor="Quan, Thanh Tho
and Sombattheera, Chattrakul
and Pham, Hoang-Anh
and Tran, Ngoc Thinh",
title="ViPPS: Building a Multimodal Dataset for Physics Problem Solving in Vietnamese",
booktitle="Multi-disciplinary Trends in Artificial Intelligence",
year="2026",
publisher="Springer Nature Singapore",
address="Singapore",
pages="308--319",
isbn="978-981-95-4960-3"
}