AnandforU/youtube-comment-insights-chatml

YouTube Comment Insights - ChatML

Overview

This dataset contains instruction-tuning samples for structured YouTube comment analysis.

The dataset is formatted in ChatML conversational format and is intended for supervised fine-tuning (SFT), QLoRA, and instruction tuning of large language models.

Each sample contains:

• sentiment
• tone
• pros
• cons

---

Dataset Statistics

• ~20k training samples
• ~2k validation samples
• Multilingual YouTube comments
• Structured JSON assistant outputs

---

Files

File	Description
train_chatml.jsonl	Training split in ChatML format
val_chatml.jsonl	Validation split in ChatML format

---

ChatML Format

Each sample follows this conversational structure:

{
  "messages": [
    {
      "role": "system",
      "content": "Analyze the YouTube comment and return ONLY valid JSON."
    },
    {
      "role": "user",
      "content": "Amazing editing but terrible audio quality."
    },
    {
      "role": "assistant",
      "content": "{\"sentiment\":\"Negative\",\"tone\":\"Frustrated\",\"pros\":[\"editing\"],\"cons\":[\"audio quality\"]}"
    }
  ]
}

---

Assistant Output Schema

{
  "sentiment": "Positive | Negative | Neutral",
  "tone": "single short emotion",
  "pros": [strings],
  "cons": [strings]
}

---

Data Generation Pipeline

1. Raw YouTube comments were cleaned and filtered.

2. Existing ground-truth sentiment labels were preserved.

3. GPT-5-Nano was used to generate:

   • tone
   • pros
   • cons

4. Samples were converted into ChatML instruction-tuning format.

---

Intended Use

This dataset is intended for:

• supervised fine-tuning (SFT)
• QLoRA training
• instruction tuning
• creator analytics research
• structured sentiment extraction

---

Notes

• Sentiment labels are ground-truth dataset labels.
• Tone, pros, and cons were synthetically generated using GPT-5-Nano.
• The dataset intentionally retains multilingual and code-switched YouTube comments to better reflect real-world creator audiences.