BSC-LT/ALIA-2606-DPO-helpfulness

Dataset Card for BSC Multilingual Synthetic Helpfulness Preferences

Dataset Summary

This dataset consists of synthetic helpfulness preference data generated to align language models across five languages: Catalan, Spanish, English, Basque, and Galician.

Building on the PKU-SafeRLHF and Tulu 3/Ultrafeedback methodologies for creating preference data, this dataset leverages an LLM-as-a-judge approach to automatically score and pair model responses to a massive pool of helpfulness prompts.

Dataset Details

Generation and Curation Process

1. Instruction Gathering: Prompts were sampled from the ALIA SFT mixture, BSC-LT/ALIA-2606-SFT.
2. Response Generation: For each prompt, five models were randomly selected from a pool of intermediate checkpoints within the ALIA and Salamandra families to generate alternative responses.
3. Helpfulness Judging: DeepSeek-V3-0324 acted as the helpfulness judge. Given a generic definition of helpfulness together with a set of examples, the judge assigned each response a score from 1 (least helpful) to 5 (most helpful).

Heuristic Filtering for Preference Pairs

To generate high-quality preference pairs, the following filtering rules were applied:

• Only responses with ratings of 1–3 were eligible as rejected responses.
• Only responses with a rating of 5 were eligible as chosen responses.
• For each prompt, the lowest-scoring eligible response was selected as the rejected response, while the highest-scoring eligible response was selected as the chosen response.

This pipeline yielded approximately 52k high-quality helpfulness preference pairs with strong separation between the chosen and rejected responses.

Dataset Size and Distribution

The table below presents the distribution of helpfulness prompts by language, detailing the number of examples contributed from each language:

Language	Instances
ca	7535
es	8836
en	28120
eu	4118
gl	3472
Total	52081

---

Dataset Structure

Data Fields

Each item in this dataset contains the following fields:

• id: Unique identifier for the item.
• lang: Language of the prompt.
• prompt: The adversarial attack user prompt.
• chosen: The chosen safety response (scored 0 or -1).
• rejected: The rejected unsafe response (scored ≤ -3).
• type: Type of preference data (always set to "helpfulness" for this dataset).
• metadata: Additional information regarding the models that generated the prompt, the chosen/rejected responses, and their exact helpfulness scores.
• preference_origin: The origin of the labels (always set to "synthetic" due to the LLM-as-a-judge scoring methodology).

Data Instances

An example of a dataset instance looks as follows:

{
  "id": "bsc_uf_v2.0_aya-dataset_1321",
  "lang": "en",
  "prompt": [
    {
      "role": "user",
      "content": "How do neural networks learn and make decisions?"
    }
  ],
  "chosen": [
    {
      "role": "assistant",
      "content": "Neural networks learn by adjusting their internal parameters, called weights, through a process called training. During training, the network processes input data and compares its predictions to the actual target values. The difference between the predicted and actual values is calculated using a loss function, which measures how far off the predictions are.\n\nThe network then uses an optimization algorithm, such as stochastic gradient descent, to update the weights in a way that minimizes the loss. This involves calculating the gradient of the loss function with respect to each weight and moving the weight in the opposite direction of the gradient. The magnitude of the update is determined by the learning rate, a hyperparameter that controls how large the steps are during optimization.\n\nOnce trained, the network can make decisions by processing new input data and producing predictions based on the learned weights. The confidence in these predictions depends on the quality of the training data, the complexity of the network, and the amount of training performed. Well-trained networks with sufficient data and appropriate architecture can make accurate and confident decisions, while poorly trained networks may produce unreliable or uncertain predictions."
    }
  ],
  "rejected": [
    {
      "role": "assistant",
      "content": "Neural networks learn by adjusting the weights of their connections based on training data. They make decisions through a process of input processing, activation, and output generation. The confidence score for this answer is 99%."
    }
  ],
  "type": "helpfulness",
  "metadata": {
    "chosen_score": 5.0,
    "rejected_score": 3.0,
    "chosen_origin": "Salamandra-40b_pre-1.4_sft-4.05_ali-2.1",
    "rejected_origin": "Salamandra-40b_pre-1.4_sft-4.0_ali-2.1",
    "prompt_origin": "mix_helpfulness",
    "prompt_id": "aya-dataset_1321",
    "dataset_name": "bsc_uf_v2.0"
  },
  "preference_origin": "synthetic"
}

Data Origin

Prompts were sampled from the ALIA SFT mixture, BSC-LT/ALIA-2606-SFT.

Additional Information

Creators

Language Modeling and Annotation Teams, AI Institute, the Barcelona Supercomputing Center (ai_institute_languagemodeling@bsc.es) (ai_institute_annotators@bsc.es).

Funding

This work is funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project Desarrollo de Modelos ALIA.

This work has been promoted and supported by the Government of Catalonia through the Aina Project.

Licensing Information

This work is licensed under CC-BY 4.0.

Acknowledgements

We acknowledge EuroHPC Joint Undertaking for awarding the project ID EHPC-AI-2024A05-046 access to MareNostrum5 at BSC, Spain.

Citation Information

WIP