This dataset was cleaned and reorganized to improve readability and make it more suitable for sharing and downstream processing.
The main changes are as follows:
• Overall organization:
Adopted a clear column order based on three logical groups:
• Identification
• Category / Topic and Upload Time
• Description:
Then performed the following
• Column renaming (for clarity and consistency):
• youtube.title → title
• content_parent_category → parent_category
• content_fine_category → fine_category
• content_metadata.description → description
• content_metadata.fps → fps
• text_to_speech → transcript_text
• text_to_speech_word_count → word_count
• Column removal:
• Dropped purely technical fields: original_json_filename, original_video_filename.
• Removed youtube_age_limit -> it was all zeroes
• Dropped redundant content_metadata fields such as storylines.scenes, trimmingSuggestions, and other repeated description-like columns.
• Data normalization:
• Converted youtube_upload_date (stored as 20160511) into a proper datetime column (youtube_upload_date_dt).
• Flattened single-element lists in youtube_* columns.
• Retained complex JSON fields (qAndA, scenes, timecoded_text_to_speech) instead of fully unnesting them because they seem relevant for future engineering tasks and json keeps the data structured.
• Final structure:
Reordered columns to highlight the most relevant, human-readable fields first:
video_id, title, upload_date, parent_category, fine_category, description, duration_seconds, resolution, fps, transcript_text, word_count, followed by all remaining metadata columns.