README.md

metadata

license: mit
task_categories:
  - text-generation
language:
  - si
  - en
tags:
  - recommendation-system
  - emotions
  - sinhala
  - singlish
size_categories:
  - 1K<n<10K

emoBooks Dataset 📚✨

Overview

The emoBooks dataset is a curated collection of conversational samples designed to train AI models for emotion-aware book recommendations. It focuses on understanding user emotions and providing relevant book suggestions in "Singlish" (transliterated Sinhala).

The dataset follows a Match/Switch logic:

• Match: Recommend books that align with the user's current emotional state.
• Switch: Recommend books that help transition the user to a more positive emotional state (e.g., from Sadness to Joy).

Dataset Structure

The dataset is provided in a standard ChatML/OpenAI format, making it ideal for instruction fine-tuning of models like Llama-3, Mistral, or Gemma.

Format

Each entry is a JSON object with a messages list:

{
  "messages": [
    { "role": "user", "content": "I feel very lonely and sad today." },
    { "role": "assistant", "content": "I can sense you're feeling that way. Would you like a book that matches your current mood, or something to help you switch to a more positive state?" },
    { "role": "user", "content": "I'd like something that matches how I feel." },
    { "role": "assistant", "content": "I understand. Since you're looking for something that matches your mood, I recommend 'Maha Nidagala Saha Abirahasa' by 'Dhigala'. It's a deep 'Translations' book that resonates with those feelings." }
  ]
}

Emotion Categories

The dataset covers 8 primary emotional states:

1. Sadness (0)
2. Joy (1)
3. Love (2)
4. Anger (3)
5. Fear (4)
6. Surprise (5)
7. Disgust (6)
8. Calm (7)

Book Categories

The recommendations include various genres:

• Children's Stories (ළමා කතා)
• Novels (නවකතා)
• Translations (පරිවර්තන)
• Biographies (චරිතාපදාන)
• Short Stories (කෙටි කතා)

Data Generation

This dataset was synthetically generated using a custom pipeline that:

1. Transliterated Sinhala book titles and authors into Singlish.
2. Mapped book categories to primary emotions.
3. Generated conversational templates for various user-assistant interactions.

Intended Use

This dataset is intended for:

• Fine-tuning Large Language Models (LLMs) for recommendation systems.
• Research in emotion-aware conversational AI.
• Developing bilingual (Sinhala/English) literary assistants.

Licensing

[MIT License] - Free to use for research and development.