pdelobelle/nemotron-dutch-mt

Nemotron Post-Training Dataset (Dutch Translation)

Machine-translated Dutch version of NVIDIA's Nemotron Post-Training Dataset, specifically the chat conversations.

Dataset Details

• —Source: `nvidia/Nemotron-Post-Training-Dataset-v2` (chat split)
• —Translation: English → Dutch using `Unbabel/Tower-Plus-9B`
• —Size: 445,287 conversations with 1,327,548 total messages
• —Format: Conversational data with original structure preserved

Dataset Statistics

• —Total conversations: 445,287
• —Total messages: 1,327,548
• —Average messages per conversation: 2.98
• —Message role distribution:
• —System: 444,567 messages
• —User: 443,054 messages
• —Assistant: 439,927 messages

Schema

Each conversation contains:

• —uuid: Unique conversation identifier from original dataset
• —category: Content category (all "chat" for this split)
• —version: Dataset version ("v2")
• —messages: Array of message objects with:
• —role: Message role (system/user/assistant)
• —content: Machine-translated Dutch message content
• —translation_metadata: Translation process metadata including job IDs and durations

Quality Notice

⚠️ This is machine-translated content. Translation quality varies and has not been manually reviewed. While the Tower-Plus-9B model provides high-quality translations, some artifacts or inconsistencies may remain. The original thinking tokens (<think></think>) have been preserved and cleaned from any Dutch translation artifacts.

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("pdelobelle/nemotron-translated-conversations")

# Access a conversation
conversation = dataset['train'][0]
print(f"Conversation {conversation['uuid']} has {len(conversation['messages'])} messages")

# Iterate through messages
for message in conversation['messages']:
    print(f"{message['role']}: {message['content']}")

Translation Process

The translation was performed by:

1. 1.Extracting individual messages from the original conversation format
2. 2.Translating each message using the Tower-Plus-9B model
3. 3.Reconstructing conversations while preserving original metadata
4. 4.Cleaning translation artifacts (Dutch <denk> tokens replaced with English <think>)

License

This dataset inherits the ODC-By license from the original Nemotron dataset. You are free to:

• —Share and use the dataset for any purpose
• —Create works based on the dataset
• —Modify, transform and build upon the dataset

With the requirement to:

• —Give appropriate credit to the original NVIDIA Nemotron creators
• —Indicate if changes were made

Citation

If you use this dataset, please cite both the original Nemotron dataset and acknowledge the translation work:

@misc{nemotron-dutch-translation,
  title={Nemotron Post-Training Dataset (Dutch Translation)},
  author={Pieter Delobelle},
  year={2025},
  note={Dutch translation of NVIDIA Nemotron Post-Training Dataset using Tower-Plus-9B}
}

Original Nemotron dataset:

@misc{nvidia-nemotron-dataset,
  title={Nemotron Post-Training Dataset v2},
  author={NVIDIA},
  year={2024},
  url={https://huggingface.co/datasets/nvidia/Nemotron-Post-Training-Dataset-v2}
}