khopilot/khmer-education-pdf-cleaned

Khmer Education PDF - Role-Playing Teaching Methods (Cleaned)

Dataset Description

This dataset contains cleaned and Unicode-corrected text extracted from a 48-page Khmer educational PDF about role-playing teaching methods (ការបងៀនតាមវិធីសម្មែងតួ). The text has undergone comprehensive Unicode repair to fix 537 orphaned COENG characters and other extraction artifacts.

Dataset Summary

• Language: Khmer (km)
• Source: Educational PDF - "New Generation Pedagogical Research Center"
• Topic: Role-playing teaching methods and pedagogical research
• Pages: 48
• Characters: 40,806 (after cleaning)
• License: CC-BY-4.0

Supported Tasks

• Khmer language modeling
• Text generation
• Educational content analysis
• Tokenization benchmarking
• NLP research

Dataset Structure

Data Files

1. pdf_text_fixed.txt - Main cleaned text file (40,806 characters)
2. coeng_repair_comparison.txt - Before/after repair analysis
3. pdf_unicode_analysis_report.txt - Original Unicode analysis
4. FINAL_PDF_ANALYSIS_REPORT.md - Comprehensive technical report

Data Fields

The main text file (pdf_text_fixed.txt) contains:

• Format: Plain text (UTF-8 encoded)
• Encoding: Unicode NFC normalized
• Structure: Sequential page extraction with page markers

Data Cleaning Process

Unicode Fixes Applied

1. COENG Character Repair (537 fixes)

• Removed 295 COENG before vowels
• Removed 159 COENG before spaces
• Removed 32 COENG before punctuation
• Removed 23 COENG before diacritic signs
• Fixed 19 double COENG patterns
• Removed 8 other orphaned COENG

2. Validation Results

• ✅ All COENG characters now properly formed (1,753 valid clusters)
• ✅ NFC Unicode normalization verified
• ✅ No invisible characters (zero-width spaces, joiners)
• ✅ Khmer text purity: 85.1%

Quality Metrics

Metric	Before	After	Improvement
Total characters	41,343	40,806	-537
Token count (km_8k)	27,249	26,556	-2.5%
Broken tokens	13,698	13,030	-4.9%
Tokens per char	0.659	0.651	-1.2%
Avg token length	1.639	1.661	+1.3%

Source Information

Original PDF Details

• Title: អនុគរុសកាសលយជំនាន់ថ្មី (New Generation Pedagogical Research Center)
• Subject: ការបងៀនតាមវិធីសម្មែងតួ (Role-playing teaching methods)
• Author: Stanislas Kowalski (translated by Khmer educators)
• Year: 2025
• Copyright: © New Generation Pedagogical Research Center
• Pages: 48

Content Overview

The document covers:

• Student participation and engagement strategies
• Vocabulary and grammar training methods
• Role-playing teaching principles
• Classroom implementation guidelines
• Feedback techniques
• Time management for role-play activities
• Scenario templates (basic, intermediate, advanced levels)
• Assessment frameworks

Limitations and Known Issues

Remaining Challenges

1. Elevated TPC: 0.651 (target: 0.40-0.45)

   • Indicates remaining text extraction quality issues
   • ~13,000 broken single-character tokens persist

2. Word Spacing: Some incorrect word boundaries remain

   • Missing spaces between words
   • Extra spaces within words

3. OCR Artifacts: Minor character substitutions may exist

Recommended Use

• ✅ Educational content analysis
• ✅ Khmer NLP research and development
• ✅ Tokenization benchmarking
• ✅ Language model pre-training (after additional cleaning)
• ⚠️ Requires further word segmentation for production use

Usage Example

Load Text

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("khopilot/khmer-education-pdf-cleaned")

# Access the text
text = dataset['train']['text'][0]
print(f"Text length: {len(text):,} characters")

Tokenization Test

import sentencepiece as spm

# Load km_8k_production tokenizer
sp = spm.SentencePieceProcessor()
sp.load("path/to/km_8k_production/tokenizer.model")

# Tokenize
tokens = sp.encode_as_pieces(text)
print(f"Token count: {len(tokens):,}")
print(f"TPC: {len(tokens) / len(text):.3f}")

Citation

If you use this dataset, please cite:

@dataset{khmer_education_pdf_cleaned_2025,
  title={Khmer Education PDF - Role-Playing Teaching Methods (Cleaned)},
  author={Kowalski, Stanislas and Khmer Pedagogical Team},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/khopilot/khmer-education-pdf-cleaned},
  note={Unicode-corrected text with 537 COENG repairs}
}

Dataset Creation

Curation Process

1. PDF text extraction using pypdf
2. Unicode analysis with custom tools
3. Automated COENG repair (537 fixes)
4. Validation with km_8k_production tokenizer
5. Quality metrics calculation

Tools Used

• Python 3.14
• pypdf 6.1.1
• sentencepiece 0.2.1
• khmer-data-curation-toolkit v1.0.0

Processing Scripts

Available in the repository:

• analyze_pdf_unicode.py - Unicode analysis
• fix_pdf_coeng.py - COENG repair
• See FINAL_PDF_ANALYSIS_REPORT.md for full methodology

License

Text Content: CC-BY-4.0 (attributed to original authors) Processing Scripts: MIT License

Original copyright: © 2025 New Generation Pedagogical Research Center

Contact

• Curator: khopilot
• Repository: https://github.com/khopilot/khmer-data-curation-toolkit
• Issues: Report data quality issues via HuggingFace discussions

Changelog

Version 1.0 (2025-10-13)

• Initial release
• Fixed 537 orphaned COENG characters
• Added comprehensive documentation
• Validated with km_8k_production tokenizer

---

Last Updated: October 13, 2025 Dataset Version: 1.0 Status: Stable