daruokta/t5gemma2-indonesia-chat-formatted

T5Gemma-2 Indonesian Chat & QA Dataset

A high-quality Indonesian language multi-turn conversation and reading comprehension dataset, specifically formatted for instruction tuning of sequence-to-sequence (Seq2Seq) models like T5-Gemma / T5-Gemma-2.

Dataset Description

This dataset contains over 7,400 multi-turn conversations and document-based Q&A in Bahasa Indonesia. It covers diverse topics including everyday life, technology, general knowledge, and structured document analysis. The dataset is pre-formatted into standard OpenAI/ChatML message lists, making it instantly ready for fine-tuning.

📚 Data Provenance (Sumber Data)

Konfigurasi indoqa_documents dan indoqa_sft pada dataset ini bersumber dan dibangun menggunakan dataset jakartaresearch/indoqa sebagai bahan dasar utamanya, yang kemudian diproses dan diformat ulang agar kompatibel dengan pelatihan instruksi (instruction tuning) model sequence-to-sequence (Seq2Seq).

📢 Updates

• —July 2026 Update: Added the chat_orpo preference configuration containing 1,000 high-quality prompt-preference pairs with annotated rationales and flaws, specifically designed for alignment optimization (ORPO/DPO).
• —June 2026 Update: Added input_tokens and target_tokens fields to the SFT configurations (chat_sft and indoqa_sft) representing the pre-computed token lengths calculated using the google/t5gemma-2-270m-270m tokenizer.
• —June 2026 Update: The chat dataset has been comprehensively revamped! It now contains 3,000 high-quality multi-turn conversations. We added 500 new Agentic Prefix-Task conversations, and the existing 2,500 conversations were completely reformatted and converted to match this new unified standard.

Dataset Structure

The repository is divided into five distinct configurations:

1. 1.`chat_multiturn`: 3,000 curated multi-turn conversational data (500 new Agentic Prefix-Task + 2,500 reformatted existing data). Each entry contains a single column messages which is a standard ChatML list of message objects (role, content).
2. 2.`indoqa_documents`: ~4,400 reading comprehension and factual Q&A examples derived from Indonesian documents (merged train & test sets) using jakartaresearch/indoqa as its foundation. Each entry contains a single column messages which is a standard ChatML list.
3. 3.`chat_orpo`: 1,000 preference alignment pairs designed for ORPO/DPO training. Each entry contains:
4. 4.id: Unique identifier.
5. 5.prompt: The user query/instruction context.
6. 6.chosen: The high-quality/preferred response.
7. 7.rejected: The low-quality/rejected response (containing common failures like repetition, formatting issues, or hallucination).
8. 8.flaw: Short explanation of the issue present in the rejected response.
9. 9.rationale: Explanation of why the chosen response is preferred.
10. 10.`chat_sft`: Flattened and pre-processed SFT format of the chat_multiturn dataset. Contains input (full conversation context with formatting tokens), target (the expected assistant response), and pre-computed token lengths.
11. 11.`indoqa_sft`: Flattened and pre-processed SFT format of the indoqa_documents dataset. Contains input, target, and pre-computed token lengths.

Usage

You can easily load a specific subset using the Hugging Face datasets library:

from datasets import load_dataset

# 1. Load the multi-turn conversational dataset
ds_chat = load_dataset("daruokta/t5gemma2-indonesia-chat-formatted", "chat_multiturn")
print(ds_chat['train'][0]['messages'])

# 2. Load the IndoQA reading comprehension dataset
ds_qa = load_dataset("daruokta/t5gemma2-indonesia-chat-formatted", "indoqa_documents")
print(ds_qa['train'][0]['messages'])

# 3. Load the ORPO preference alignment dataset
ds_orpo = load_dataset("daruokta/t5gemma2-indonesia-chat-formatted", "chat_orpo")
print(ds_orpo['train'][0])

# 4. Load pre-processed SFT configurations (e.g., chat_sft)
ds_chat_sft = load_dataset("daruokta/t5gemma2-indonesia-chat-formatted", "chat_sft")
print(ds_chat_sft['train'][0])

Statistics

• —Total examples: ~13,800 across all subsets (including SFT splits)
• —Language: Indonesian (Bahasa Indonesia)
• —Subsets: chat_multiturn, indoqa_documents, chat_sft, indoqa_sft, chat_orpo
• —Formats: Standard ChatML format for raw documents; input/target format for SFT splits; prompt/chosen/rejected format for ORPO splits.
• —Created: May - July 2026