metadata

license: other
task_categories:
  - question-answering
  - text-generation
language:
  - id
size_categories:
  - 100K<n<1M

Indo-Instruct-v1

Indo-Instruct-v1 is a high-quality Indonesian instruction-tuning dataset consisting of 488,910 single-turn examples. The dataset is primarily sourced from community contributions and various public datasets, and it aims to support the development of instruction-following language models in Bahasa Indonesia.

Dataset Features

Language: Indonesian (Bahasa Indonesia)
Number of Samples: 488,910

Format: Follows a messages structure compatible with many open-source LLM training frameworks:

{
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ]
}

Style:
- All examples are single-turn (no multi-turn dialogue).
- Prompts are typically written in formal or rigid Indonesian, with some using standard, even textbook-like phrasing.

Data Quality

The dataset has undergone extensive deduplication and curation processes, with approximately 250,000 samples removed during quality filtering to ensure higher standards. Despite this rigorous preprocessing, given the substantial remaining volume (488,910 examples), some imperfections may still be present. Users should expect:

Generally high quality instruction-response pairs suitable for training
Occasional inconsistencies in formatting, style, or content quality that escaped automated filtering
Recommended additional preprocessing for production use cases requiring strict quality standards

The significant reduction in dataset size demonstrates our commitment to quality over quantity, though we encourage users to apply domain-specific filtering based on their requirements.