Tralalabs/vietnamese-youtube-dataset

Vietnamese YouTube Multi-Video Commentary Dataset 🚀

A structured, clean text corpus compiling video metadata and public audience commentary extracted from the Vietnamese YouTube and YouTube Shorts ecosystem. This dataset is optimized for natural language processing (NLP), linguistic analysis, sentiment classification, and machine learning model fine-tuning. 📈

The entire corpus is consolidated into a single high-performance Apache Parquet (.parquet) file, allowing for compressed storage, fast columnar reads, and direct integration with modern AI pipelines. 📦

---

🔎 Dataset Overview

• Target Language: Vietnamese (Tiếng Việt) including contemporary online colloquialisms, abbreviations, and localized emoji expressions. This dataset may include English comments from Vietnamese videos, but the comments are almost all Vietnamese. 🇻🇳
• Source Platform: YouTube and YouTube Shorts (covering highly engaged cultural comparisons, viral media discussions, and gaming communities). 🎥
• Storage Format: Apache Parquet (.parquet) for optimized memory mapping and out-of-the-box compatibility with frameworks like pandas, Dask, and Hugging Face Datasets. 💾
• Structural Design: Multi-video indexing strategy that groups a singular validated video title metadata row alongside thousands of its corresponding community comments under a uniform tracking identifier. 📂

---

📊 Database Schema & Layout

To ensure frictionless scaling across large quantities of unique media entries, the corpus follows a strict relational flat-file schema:

Column Name	Data Type	Key Role	Description	Example Value
video_id	string	Primary Key / Partition	The unique 11-character YouTube alphanumeric video identifier string.	"e.g, iI4UuaHx9lA"
type	string	Categorical Filter	Declares the row content type (title or comment).	"e.g, comment"
text	string	Content Payload	The raw string payload extracted directly from the platform.	"e.g, Chúng ta có vật lý cocomelon trước khi gta 6 ra mắt..."
collected_at	timestamp	Temporal Metadata	ISO 8601 timestamp tracking when the content stream was parsed.	"e.g, 2026-06-13T17:00:00Z"

🛠️ Data Sequencing Rules Per Video:

1. The Title Anchor: The initial row for any unique video_id contains type: "title", saving the official video title text to preserve semantic context.
2. The Comment Stream: All remaining rows allocated to that video_id contains type: "comment", storing the parsed text string of a user comment.

---

💻 Loading and Filtering the Dataset

Using standard Python data processing packages, queries, slices, and statistical aggregates can be run over millions of rows cleanly without exhausting system RAM. ⚡

🐍 Ingesting the Master File via Pandas:

import pandas as pd

# Load the entire multi-video dataset into memory
df = pd.read_parquet("vietnamese_youtube_dataset.parquet")

# Inspect database dimensions and structural integrity
print(f"Total records in corpus: {len(df)}")
print(df.info())

🔍 Isolating Content for a Specific Video:

# Extract all records corresponding to a designated video identifier
target_video_id = "iI4UuaHx9lA"
video_subset = df[df["video_id"] == target_video_id]

# Extract the video title string and isolate the comment rows
video_title = video_subset[video_subset["type"] == "title"]["text"].values[0]
comments_only = video_subset[video_subset["type"] == "comment"]

print(f"Video Title: '{video_title}'")
print(f"Total Parsed Comments: {len(comments_only)}")

📊 Aggregating Content Counts:

# Calculate the exact number of comment entries available per video index
comment_distribution = df[df["type"] == "comment"].groupby("video_id").size()
print(comment_distribution)

---

⚙️ Scraping & Data Pipeline Integration Template

To seamlessly expand this dataset with additional videos using stream-based python extractors, utilize the following robust programmatic loop structure:

import pandas as pd
from youtube_comment_downloader import YoutubeCommentDownloader
from itertools import islice
import datetime

def fetch_and_format_video_data(video_url, custom_title, max_comments=2599):
    # Parse video_id cleanly out of standard watch or shorts links
    if "=" in video_url:
        video_id = video_url.split("v="[-1].split("&")[0]
    else:
        video_id = video_url.split("/shorts/")[-1].split("/")[-1].split("?")[0]
        
    downloader = YoutubeCommentDownloader()
    timestamp_str = datetime.datetime.utcnow().isoformat() + "Z"
    
    # Initialize the required single metadata title row
    buffered_rows = [{
        "video_id": video_id,
        "type": "title",
        "text": custom_title,
        "collected_at": timestamp_str
    }]
    
    try:
        # Request stream generator from endpoint
        raw_stream = downloader.get_comments(video_id)
        bounded_stream = islice(raw_stream, max_comments)
        
        for record in bounded_stream:
            text_payload = record.get('text', '').strip()
            if text_payload:
                buffered_rows.append({
                    "video_id": video_id,
                    "type": "comment",
                    "text": text_payload,
                    "collected_at": timestamp_str
                })
    except Exception as error:
        print(f"Execution Error encountered while parsing stream for {video_id}: {error}")
        
    return pd.DataFrame(buffered_rows)

---

🧠 Recommended Natural Language Processing Applications

• Colloquial Dialectology Mapping: Track contemporary morphological shifts, online shorthand variances, and phonetic substitutions inside the modern Vietnamese text ecosystem. 📑
• Large Language Model (LLM) Fine-Tuning: Provide natural, conversational language variants for target alignment steps or reinforcement training loops. 🤖
• Unsupervised Sentiment Clustering: Extract word embeddings from the text payload column to build topical, multi-axis clustering layers on community engagement patterns. 🎯

---

📜 Dataset Licensing & Distribution

This collaborative corpus is distributed globally under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license framework. ⚖️

• Attribution: Any usage, derivation, re-distribution, or academic reference of these text objects must properly cite the dataset pipeline. ✍️
• Non-Commercial: The dataset payload may not be bundled, modified, or re-packaged for commercial sale or commercial monetization vectors. It remains open and uninhibited for public academic and independent developers. 🛡️