lopanapol/sample-thai

Cleaned Text Dataset

This repository contains a sample cleaned text dataset, originally sourced from JSONL data and processed for enhanced quality and utility in various Natural Language Processing (NLP) tasks. The data is stored in optimized formats (e.g., Parquet) on the Hugging Face Hub. The primary goal of this cleaning process is to remove noise, standardize formatting, and improve the overall signal-to-noise ratio, making the data more suitable for training, fine-tuning, and evaluating machine learning models.

Source Data

The raw data for this cleaned dataset originated from the Hugging Face dataset: ChavyvAkvar/fineweb-2-1M-Sample-Thai

Cleaning Methodology

The cleaning process is implemented in minimal.py within the clean_text_data function. It applies the following transformations to the text field of each JSONL record:

1. HTML Tag Removal: All HTML or XML tags (e.g., <p>, <div>, <a href="...">) are identified and removed from the text. This prevents structural web elements from interfering with text-based analysis.

2. URL Removal: URLs (starting with http:// or https://) are removed. This step is crucial for datasets where URLs are not relevant content and might introduce noise or privacy concerns.

3. Whitespace Normalization:

   • Newline characters (\n, \r) are replaced with single spaces.
   • Multiple consecutive spaces are reduced to a single space.
   • Leading and trailing whitespace from the entire text string are trimmed.

4. Non-Printable Character Filtering: Any non-printable Unicode characters are removed. This ensures that the text only contains characters that are typically rendered and understood by humans and NLP models.

5. Removal of Unicode Replacement Character (U+FFFD): The Unicode Replacement Character (U+FFFD, often appearing as ``) is explicitly removed. This character typically indicates that text could not be decoded properly and can make output appear garbled.

6. Keyword Filtering (Optional) - Enhanced with pythainlp Tokenization: If a keyword is provided during the cleaning process, only records whose text field contains this keyword will be included in the output. Critically, for Thai language text, pythainlp's word tokenizer is now utilized.

Why these cleaning steps?

• Improved Model Performance: Clean data directly leads to better model performance by reducing irrelevant features and focusing the model on meaningful linguistic patterns.
• Reduced Noise: Removing HTML, URLs, and extraneous whitespace eliminates common sources of noise in web-scraped or uncurated text data.
• Standardized Format: Consistent formatting makes the data easier to process and integrate into various NLP pipelines without needing custom preprocessing scripts for each use case.
• Enhanced Readability: The cleaned text is generally more readable and comprehensible for human review.

Cleaned Data Format

The output of the cleaning process initially produces a JSONL file. The dataset's records, however, are stored in optimized formats (like Parquet) on the Hugging Face Hub, and each record adheres to the following enriched structure: This expanded format aims to provide more metadata for better filtering, analysis, and management of the dataset.

{
  "id": "...",                   // (string) Unique identifier for the record, stripped of "<urn:uuid:>" prefix.
  "text": "...",                 // (string) The cleaned text content.
  "text_hash": "...",            // (string) SHA256 hash of the 'text' content, useful for deduplication and integrity checks.
  "text_length": ...,            // (integer) Character length of the 'text' content.
  "url": "...",                  // (string) The original URL of the content.
  "domain": "...",               // (string) The extracted domain from the 'url' (e.g., "example.com").
  "url_path": "...",             // (string) The extracted path from the 'url' (e.g., "/news/article-title").
  "date": "...",                 // (string) The date associated with the original data collection (ISO 8601).
  "language": "...",             // (string) The detected language code (e.g., "tha").
  "cleaned_timestamp": "...",     // (string) ISO 8601 timestamp (UTC) indicating when the cleaning was performed.
  "source_name": "...",          // (string, optional) Human-readable name of the original data source (e.g., "FineWeb", "CommonCrawl").
  "content_type": "...",         // (string, optional) Type of content (e.g., "article", "blog_post", "forum_comment").
  "language_confidence": ...     // (float, optional) Confidence score if language detection is automated.
  "keyword": "..."                 // (string, optional) The keyword used to filter the text, if applicable.
}

Rationale for New Fields:

• id (plain UUID): Stripping the URN prefix makes the ID cleaner and more directly usable as a unique identifier across various systems (e.g., database primary keys).
• text_hash: Enables efficient detection of duplicate content and verification of data integrity.
• text_length: Provides a quick metric for content size, useful for filtering or analysis without reloading the full text.
• domain & url_path: Facilitates grouping and analysis of content by source website or specific sections of a website, enhancing data queryability.
• source_name, content_type, language_confidence: These optional fields serve as placeholders for further data enrichment, allowing for more granular filtering, source attribution, and quality assessment if such metadata becomes available.
• keyword: When a keyword filter is applied, this field explicitly records the keyword used, enabling easier post-processing, analysis, and interpretation of the filtered dataset.

Keyword Distribution

The following tables show the distribution and counts of keywords found in the dataset, separated by single and multiple keyword entries.

Single Keywords

Keyword	Translation	Count
กระเทียม	Garlic	246
ท่องเที่ยว	Travel/Tourism	19465
นิยาย	Novel/Fiction	7026
ประวัติศาสตร์	History	2240
วรรณกรรม	Literature	2090
อาหาร	Food	20812
งานวิจัย	Research	10298
บทสัมภาษณ์	Interview	2317
การ์ตูน	Cartoon	12482
dopamine	Dopamine	67
การปลูกต้นไม้	Tree Planting	885
อาชีพเสริม	Side Job	1478
street food	Street Food	359
ขนมไทยโบราณ	Traditional Thai desserts	206
ragnarok online	Ragnarok Online	105
ย้อนยุค	Retro	3300
vegan	Vegan	235
รายการทีวี	TV shows	553
ภาษาซี	C Language	59

Multiple Keywords

Keyword	Translation	Count
ปรัชญา, ศาสนา, บทความ, ผู้เขียน	Philosophy, Religion, Article, Author	24

Potential Use Cases

This cleaned data is ideal for:

• Language Model Pretraining: Providing a cleaner corpus for foundational model training.
• Fine-tuning: Enhancing the quality of fine-tuning datasets for specific tasks like summarization, question answering, or text generation.
• Text Classification: Improving feature extraction for classification tasks.
• Information Retrieval: Reducing noise for better search results.
• Data Analysis: Facilitating more accurate and insightful textual analysis.

Contribution and Feedback

Contributions are welcome! If you have suggestions for additional cleaning steps, improvements to the existing methodology, or encounter any issues, please open an issue or pull request.

Support