krotreaksmey/khmer-math-textbook_v2

Khmer Math & STEM Textbook Line-Level OCR Dataset

Benchmark Dataset for Cambodian Secondary STEM, Advanced Calculus & National Examinations (101,463 Samples)

An authoritative, benchmark-grade dataset of 112,113 line-level image crops and verified ground-truth transcriptions extracted from official Cambodian Ministry of Education, Youth and Sport (MoEYS) STEM textbooks, national examination series, and high school competition archives.

Designed specifically for training and benchmarking state-of-the-art Optical Character Recognition (OCR), Mathematical Expression Recognition (MER), and Vision-Language Models (VLM) on mixed Khmer Unicode script and complex LaTeX mathematical notation.

Major v3.0 Expansion (+30,020 New Samples — Advanced Calculus Track): This update pushes the dataset beyond the 100,000 threshold (101,463 total samples) by incorporating a comprehensive, curriculum-aligned corpus dedicated to Upper-Secondary Advanced Calculus & Analysis:

1. (Continuity of Functions): Pointwise continuity, interval continuity, piecewise functions, and Intermediate Value Theorem .
2. (Derivatives & Differentiation): Product/Quotient/Chain rules, polynomial/trigonometric/exponential/logarithmic derivatives, tangent lines, inflection points, and extreme values.
3. (Integrals & Primitives): Indefinite primitives, definite integrals, integration by parts ($\int u dv = uv - \int v du$), substitution methods, area under curves, and volumes of solids of revolution.
4. (Differential Equations): First-order linear differential equations ($y' + ay = 0$), second-order linear differential equations ($y'' + ay' + by = 0$), characteristic equations, and initial value problems (លក្ខខណ្ឌដើម).

---

Dataset Overview

Metric	Specification
Total Labeled Line Crops (Parquet)	112,113 samples in data/train-*.parquet (14 partitions, ~1.79 GB)
Total Curated Annotations (Metadata)	112,113 entries in labeled_metadata.csv & label_checkpoint.json
Languages	Khmer (km), English / Mathematical LaTeX (en)
Modality	Line-crop Images (image) + Dual-script Transcriptions (label)
Mathematical Notation	Standard LaTeX syntax enclosed in $ ... $ (\frac, \sqrt, \int, \lim, y'', \vec{u}, etc.)
Orthography Standard	Official Samdech Chuon Nath Khmer Dictionary & MoEYS Curricular Standards
Quality Assurance	Multi-pass human verification, AST syntax linting, Thai character sanitization
License	Apache-2.0

---

Comprehensive Curriculum & Grade Coverage

This dataset spans primary to upper-secondary education, covering all core STEM subjects taught across Cambodia:

Grade / Academic Level	Curricular Track	Samples (Line Crops)	Core Topics Covered
Grade 12 Advanced Calculus (NEW Expansion)	Advanced Science Track	30,020	ភាពជាប់នៃអនុគមន៍ (Continuity): Pointwise/interval continuity, limits at points of discontinuity, Intermediate Value Theorem. ដេរីវេ (Derivatives): Higher-order derivatives ($f'', y^{(n)}$), composite function differentiation, optimization, curve sketching. អាំងតេក្រាល (Integrals): Primitives, definite integrals, integration by parts, rational fraction integration, area/volume integrals. សមីការឌីផេរ៉ង់ស្យែល (Differential Equations): Linear 1st & 2nd order differential equations with constant coefficients.
Grade 12 Mathematics (Core)	Both Advanced & Basic Tracks	15,289	Advanced Track: Integrals, Limits, Complex Numbers, Differential Equations, 3D Analytic Geometry, Probability. Basic Track: Sequences, Financial Mathematics, Descriptive Statistics, Exponential Functions, Permutations & Combinations.
Grade 11 Mathematics	Both Advanced & Basic Tracks	7,820	Advanced Track: Derivatives, Numerical Sequences, Trigonometric Equations & Identities, 3D Vectors, Spatial Coordinates. Basic Track: Polynomial Operations, Linear Programming, Quadratic Inequalities, Basic Trigonometric Ratios.
Grade 10 Mathematics	Both Advanced & Basic Tracks	6,121	Advanced Track: Set Theory, Advanced Polynomial Factoring, Quadratic Systems, Analytic Geometry, In-depth Trigonometry. Basic Track: Linear Equations & Inequalities, Real Number Properties, Quadratic Equations, Right-Triangle Geometry.
Complex Numbers & Limits (Lim Phalkun Collection)	Upper Secondary & Advanced Exam Prep	3,327	Specialized advanced mathematics on complex numbers and limits (គណិតវិទ្យា កុំផ្លិច និងលីមីត រៀបរៀងដោយលោកគ្រូ លឹម ផល្គុន) featuring algebraic and trigonometric forms of complex numbers, De Moivre's theorem, indeterminate limits ($\frac{0}{0}, \frac{\infty}{\infty}, \infty - \infty, 0 \cdot \infty$), trigonometric limits, and detailed step-by-step problem solutions.
Grade 9 National Diploma (Bac I) 1981–2012 Exam Archives	31-Year National Examination Series	1,857	Complete historical collection of lower-secondary diploma exam papers and solutions (កម្រងគ្រប់វិញ្ញាសា និងចម្លើយ ប្រឡងសញ្ញាបត្រមធ្យមសិក្សាបឋមភូមិ ១៩៨១–២០១២) covering algebra, geometry proofs, polynomial divisions, radical simplifications, and equation systems.
Grade 9 Math Diploma Exam Prep Collection	National Diploma Examination Series	2,530	Dedicated national Grade 9 diploma examination archives (កម្រងវិញ្ញាសាគណិតវិទ្យាសម្រាប់ត្រៀមប្រឡងឌីប្លូម) featuring multi-part exam problems, linear equation systems, geometric proofs, and formal solution walkthroughs.
Grade 9 MoEYS Mathematics Textbook	Standard Middle School Curriculum	4,317	Standard Grade 9 textbook chapters covering algebra, proportions, functions, basic statistics, circles, and Pythagorean theorem applications. (Total Grade 9 Math: 8,704 crops)
Grades 6–8 Foundation Mathematics	Middle School Foundation	6,349	Fractions, Decimals, Integer Arithmetic, Angles, Polygons, Perimeter & Area, Ratios and Percentages, Pre-Algebra.
Physics Grades 11–12	Pure Science Track	6,531	Classical Mechanics (Newton's Laws, Kinematics), Thermodynamics, DC & AC Circuits, Electromagnetism, Wave Optics, Nuclear Physics & Physical Constants.
Calculus & Analysis Track (Extended)	Calculus Intensive	2,686	Specialized derivatives of rational, trigonometric, exponential, and piecewise functions with complete step-by-step solutions.
Geometry & Trigonometry	Plane & Solid Geometry	1,791+	Euclidean proofs, geometric definitions, circle theorems, vectors, angle identities, and diagram annotations.
National Bac II Examination Archives	National Exam Series (2002–2024)	12,825+	Official examination papers and model answer keys from Cambodian national Grade 12 examinations spanning over two decades.

---

Mathematical Domain & Syntax Coverage

Over 75.0% of the dataset contains combined Khmer text with mathematical expressions, while 25.0% represents pure explanatory Khmer prose:

Mathematical Construct	LaTeX Syntax	Occurrences	Percentage	Example in Dataset
Equations & Systems	=, \begin{cases}	38,000+	37.5%	សមីការ $2x^2 - 5x + 3 = 0$
Fractions & Quotients	\frac{a}{b}	18,500+	18.2%	$f(x) = \frac{x^2 + 3x}{x - 1}$
Khmer Technical Terminology	សមីការ, អនុគមន៍, ដេរីវេ	16,000+	15.8%	កំណត់អនុគមន៍ $f$ លើចន្លោះ $\mathbb{R}$
Derivatives & Differential Equations	$f'(x)$, $y'' + ay' + by = 0$	12,800+	12.6%	ដេរីវេនៃ $y' = 6x^2 - 4x + 1$, $y'' - 4y' + 4y = 0$
Integrals & Primitives	\int_{a}^{b} f(x) dx	6,800+	6.7%	អាំងតេក្រាល $\int_0^1 (2x + 1) dx = [x^2 + x]_0^1$
Limits & Continuity	\lim_{x \to a}, ជាប់ត្រង់	5,400+	5.3%	$\lim_{x \to 0} \frac{\sin x}{x} = 1$, $f$ ជាប់ត្រង់ $x_0$
Greek Letters & Constants	$\alpha$, $\beta$, $\theta$, $\pi$, $\Delta$	4,600+	4.5%	ឌីសគ្រីមីណង់ $\Delta = b^2 - 4ac$
Trigonometric Functions	\sin, \cos, \tan	4,200+	4.1%	$\sin^2 x + \cos^2 x = 1$
Radicals & Square Roots	\sqrt{x}	3,800+	3.7%	ឫសការេ $\sqrt{x^2 - 4} = 3$
Inequalities & Intervals	\leq, \geq, <, >	3,500+	3.4%	វិសមភាព $2x - 3 \leq 5$
Summations & Vectors	\sum, \vec{u}	3,100+	3.1%	ផលបូកវ៉ិចទ័រ $\vec{u} + \vec{v} = \vec{0}$
Matrices & Arrays	\begin{matrix}	250+	0.2%	ម៉ាទ្រីស $\begin{pmatrix} a & b \ c & d \end{pmatrix}$

---

Dataset Format & Schema

The dataset is partitioned into high-performance Apache Parquet files (data/train-*.parquet):

krotreaksmey/khmer-math-textbook/
├── README.md                           # Comprehensive dataset documentation
├── data/
│   ├── train-00000-of-00014.parquet    # Sharded Parquet partitions with embedded images
│   ├── train-00001-of-00014.parquet
│   ├── train-00002-of-00014.parquet
│   ├── train-00003-of-00014.parquet
│   ├── train-00004-of-00014.parquet
│   ├── train-00005-of-00014.parquet
│   ├── train-00006-of-00014.parquet
│   ├── train-00007-of-00014.parquet
│   ├── train-00008-of-00014.parquet
│   ├── train-00009-of-00014.parquet
│   ├── train-00010-of-00014.parquet
│   ├── train-00011-of-00014.parquet
│   ├── train-00012-of-00014.parquet
│   └── train-00013-of-00014.parquet    # Total: 112,113 samples (~1.79 GB)
├── labeled_metadata.csv                # 101,463 entries (ID, Label)
└── label_checkpoint.json               # JSON index of ground-truth transcriptions

Schema Features

Column	Type	Description
image	datasets.Image	High-resolution grayscale or RGB crop of the text line.
label	Value("string")	Ground-truth transcription containing Khmer Unicode and LaTeX math notation.

---

Quick Start Examples

1. Load with Hugging Face datasets

from datasets import load_dataset

# Load full training split directly from Hugging Face Hub
dataset = load_dataset("krotreaksmey/khmer-math-textbook", split="train")

print(f"Total samples: {len(dataset):,}")
sample = dataset[0]

# Display sample
print("Transcription:", sample["label"])
# sample["image"].show()

2. Stream Dataset (Without Downloading Entire 1.79 GB Parquet Set)

from datasets import load_dataset

# Stream examples on-the-fly
dataset = load_dataset("krotreaksmey/khmer-math-textbook", split="train", streaming=True)

for sample in dataset.take(5):
    print("Sample:", sample["label"])

3. PyTorch DataLoader for Training OCR / Vision Models

import torch
from datasets import load_dataset
from torch.utils.data import DataLoader
from torchvision import transforms

dataset = load_dataset("krotreaksmey/khmer-math-textbook", split="train")

transform = transforms.Compose([
    transforms.Resize((48, 192)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

def collate_fn(batch):
    images = [transform(b["image"].convert("RGB")) for b in batch]
    labels = [b["label"] for b in batch]
    return torch.stack(images), labels

loader = DataLoader(dataset, batch_size=32, shuffle=True, collate_fn=collate_fn)

for batch_images, batch_labels in loader:
    print(f"Batch shape: {batch_images.shape}, First label: {batch_labels[0]}")
    break

---

Model Training Recommendations

• Pre-trained SOTA Recognition Model:
  • TinyOCRkhmer (PARSeq v4.0): Official SOTA recognition weights (best_khmer_parseq_model_101K_cer21.35%.pt, CER: 21.35%, WER: 37.67%) trained directly on this 101,463 sample dataset with full Colab training notebook.
• Architecture Compatibility:
  • Encoder-Decoder OCR: PARSeq, TrOCR, CRNN.
  • Document Vision-Language Models: Donut, Nougat, Surya, Chandra.
  • Multimodal LLMs (Fine-tuning): Qwen2.5-VL, Florence-2, PaliGemma-2.
• Recommended Image Preprocessing:
  • Resolution: 48px height x 192px width (or proportional aspect-ratio padding).
  • High vertical clarity: Khmer subscript feet (ជើងអក្សរ) and upper vowels require vertical preservation.
• Tokenization Tip:
  • Ensure your tokenizer vocabulary includes all Khmer Unicode code points (U+1780 to U+17FF) and standard LaTeX math control words (\frac, \sqrt, \int, \lim, etc.).

---

Citation & Attribution

If you use this dataset in your research or application, please cite:

@dataset{khmer_math_textbook_ocr_2026,
  author    = {krotreaksmey},
  title     = {Khmer Math & STEM Textbook Line-Level OCR Dataset (101K Samples)},
  year      = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/krotreaksmey/khmer-math-textbook}}
}

---

Acknowledgments & Contact

Curated with the Khmer OCR & AI Vision Studio. Dedicated to empowering Cambodian students, educators, and AI researchers with high-accuracy digital tools for national education.