SAWithanage/en-si-parallel-3k-llama3-format

Dataset Card for en-si-parallel-3k-llama3-format

Overview

This is a specially formatted, instruction-ready version of the original en-si-parallel-3k dataset. It has been strictly engineered to fine-tune the SAWithanage/SinLlama-Llama-3-8B-Merged base model (and other Llama-3 architectures) for English-to-Sinhala translation.

This dataset utilizes the industry-standard ShareGPT Conversational Format. By structuring the data as a standardized list of roles and contents, it allows frameworks like Unsloth, TRL, and Hugging Face to seamlessly apply native chat templates at runtime, ensuring highly stable and accurate tokenization.

Data Processing & Transformations

To optimize this dataset for base-model fine-tuning, the original dataset underwent the following programmatic transformations:

1. 1.ShareGPT Conversion: The dataset was reformatted from raw strings into a structured conversations list containing system, user, and assistant roles.
2. 2.Quote Stripping: All double quotes (") were purged from the text so the model does not artificially learn to wrap every Sinhala output in quotes.
3. 3.Randomized Shuffling: The 3,000 examples were completely shuffled to prevent catastrophic forgetting and stabilize gradient updates across different topics.
4. 4.Static System Prompt: A simple, uniform system prompt ("Translate the following English text to Sinhala.") was applied across all examples.

Row Format Example (How the Model Sees It)

Each row in the dataset contains a conversations array structured exactly like this:

[
  {
    "role": "system",
    "content": "Translate the following English text to Sinhala."
  },
  {
    "role": "user",
    "content": "For a successful sewing a loose button workflow on children's heavy school uniforms, choose a durable polyester-wrapped coreset thread line."
  },
  {
    "role": "assistant",
    "content": "ළමුන්ගේ බරැති පාසල් නිල ඇඳුම් මත ලිහිල් වූ බොත්තමක් මැසීමේ සාර්ථක ක්‍රියාවලියක් සඳහා, කල් පවතින පොලියෙස්ටර්-ඔතන ලද කෝර්සෙට් (Core-spun) නූල් රැහැනක් තෝරාගන්න."
  }
]

Original Dataset Origins

The foundational 3,000 English-Sinhala sentence pairs were synthetically generated using Gemini 3.5 Flash and adhere to a highly practical 3-tier terminology methodology:

• —Pure Sinhala for common terms.
• —Bracket bridging (Sinhala + [English]) for rare/academic terms.
• —Pure English transliteration for global standards (e.g., UI elements).