language:

• nl
  license: apache-2.0
  size_categories:
• 1K<n<10K
  task_categories:
• text-generation
  pretty_name: Orca DPO Pairs Dutch Cleaned
  dataset_info:
• config_name: default
  features:
  • name: geitje-7b-chat
    dtype: string
  • name: geitje-7b-chat_lid
    dtype: string
  • name: geitje-7b-chat_lid_prob
    dtype: float64
  • name: geitje-7b-ultra
    dtype: string
  • name: gpt-4-turbo
    dtype: string
  • name: gpt-4-turbo_lid
    dtype: string
  • name: gpt-4-turbo_lid_prob
    dtype: float64
  • name: question_dutch
    dtype: string
  • name: question_dutch_lid
    dtype: string
  • name: question_dutch_lid_prob
    dtype: float64
  • name: question_english
    dtype: string
  • name: system_dutch
    dtype: string
  • name: system_dutch_lid
    dtype: string
  • name: system_dutch_lid_prob
    dtype: float64
  • name: system_english
    dtype: string
  • name: geitje-7b-ultra_lid
    dtype: string
  • name: geitje-7b-ultra_lid_prob
    dtype: float64
    splits:
  • name: train
    num_bytes: 52416545
    num_examples: 10519
    download_size: 29554311
    dataset_size: 52416545
• config_name: dpo_all
  features:
  • name: prompt
    dtype: string
  • name: chosen
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: rejected
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
      splits:
  • name: train_prefs
    num_bytes: 46513397
    num_examples: 9467
  • name: test_prefs
    num_bytes: 5205381
    num_examples: 1052
    download_size: 28301144
    dataset_size: 51718778
• config_name: sft_gpt4_all
  features:
  • name: prompt
    dtype: string
  • name: messages
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
      splits:
  • name: train_sft
    num_bytes: 24109199
    num_examples: 9467
  • name: test_sft
    num_bytes: 2718968
    num_examples: 1052
    download_size: 14998305
    dataset_size: 26828167
    configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*
• config_name: dpo_all
  data_files:
  • split: train_prefs
    path: dpo_all/train_prefs-*
  • split: test_prefs
    path: dpo_all/test_prefs-*
• config_name: sft_gpt4_all
  data_files:
  • split: train_sft
    path: sft_gpt4_all/train_sft-*
  • split: test_sft
    path: sft_gpt4_all/test_sft-*
    tags:
• conversational
• dpo
• sft

Dataset Card for Orca DPO Pairs Dutch Cleaned

Citation

If you use this dataset, GEITje 7B Ultra (SFT) or any of its derivatives or quantizations, place cite the following paper:

@misc{vanroy2024geitje7bultraconversational,
      title={GEITje 7B Ultra: A Conversational Model for Dutch}, 
      author={Bram Vanroy},
      year={2024},
      eprint={2412.04092},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.04092}, 
}

Dataset Description

This is a new version of BramVanroy/orca_dpo_pairs_dutch, which is a dataset for preference tuning a Dutch LLM. This new version has the following differences:

• data has been deduplicated and filtered
• now has GEITje 7B Ultra responses
• includes both a split for SFT (with GPT4 responses) called sft_gpt4_all and one for preference tuning (GEITje 7B Ultra = rejected; GPT 4 Turbo = chosen) called dpo_all. Note that the SFT and DPO configs contain the same data and should probably not be used for BOTH SFT and preference tuning sequentially but only one of them. Or you can use the preference dataset with something like ORPO. For SFT, I recommend the ultra chat dataset

from datasets import load_dataset


sft_dataset = load_dataset("BramVanroy/orca_dpo_pairs_dutch_cleaned", "sft_gpt4_all")
preference_dataset = load_dataset("BramVanroy/orca_dpo_pairs_dutch_cleaned", "dpo_all")

Dataset Structure

The dataset has three configs: just the default with all data (not directly usable but here for transparency), and:

• sft_gpt4_all: for supervised finetuning, with prompt and messages (GPT4 Turbo) columns. Has splits train_sft and test_sft
• dpo_all: for preference tuning, with prompt, chosen (GPT4 Turbo), and rejected (GEITje 7B Ultra) columns. Has splits train_prefs and test_prefs

The dataset has two splits. The number of examples per split is shown as follows:

Dataset Creation

This dataset was created with this repository. The translate-hf script for translating the system and instruction columns, and answer-hf to create answers for the translated query.

So in a first step, the English first user message was translated alongside the system message with gpt-4-1106-preview to Dutch. This prompt is then fed to gpt-4-1106-preview and GEITje-7B-Ultra to generate responses, so only the user/system prompt is translated. The data is therefore single-turn.

Data filtering

On top of the automatic content filtering via the Azure API and deduplication, additional filtering was also done.

• samples with non-Latin characters are removed (very strict filtering, removes any translation tasks with non-Latin languages)
• samples with occurrences of "AI-assistent" or "AI-taalmodel" (and other derivations) are removed because these are often responses in the sense of "As an AI model, I cannot ...", which is not too useful
• samples with mentions of ChatGPT, GPT 3/4, OpenAI or ShareGPT are removed
• samples with mentions of the typical "knowledge cutoff" are removed
• samples with apologies such as "spijt me" are removed, as we are more interested in factual information and content-filled responses

Source Data

Initial Data Collection and Normalization

Initial data preparation by Intel, which in turn started from OpenOrca.

Considerations for Using the Data

Note that except for a manual analysis of around 100 random samples, this dataset has not been verified or checked for issues. However, the OpenAI filters blocked requests for translation and generation, which may indicate that some of the original prompts may contained ambiguous or potentially inappropriate content that has now been filtered.

During checking, I found that the translations of the prompt and system prompt sometimes have issues, where parts of the content was not translated. This can happen when the model interprets the input as a command, for instance, instead of translating it. This is especially true for the system messages.

Discussion of Biases

As with any machine-generated texts, users should be aware of potential biases that are included in this dataset. It is likely that biases remain in the dataset so use with caution.

Licensing Information

Since the data was created with the Azure API, and I cannot find any restrictions on how the data can be used, you can use this data however you see fit according to an apache 2.0 license.

Contributions

Thanks to Michiel Buisman of UWV for reaching out and making the creation of this dataset possible with access to Azure's API.