---
dataset_info:
  features:
  - name: prompt
    dtype: string
  - name: chosen
    list:
    - name: content
      dtype: string
    - name: role
      dtype: string
  - name: chosen_model
    dtype: string
  - name: rejected
    list:
    - name: content
      dtype: string
    - name: role
      dtype: string
  - name: rejected_model
    dtype: string
  - name: skill_chosen
    dtype: float64
  - name: skill_rejected
    dtype: float64
  - name: skill_gap
    dtype: float64
  - name: chosen_txt
    dtype: string
  - name: rejected_txt
    dtype: string
  - name: len_c
    dtype: int64
  - name: len_r
    dtype: int64
  splits:
  - name: train
    num_bytes: 88826167
    num_examples: 6884
  download_size: 44904871
  dataset_size: 88826167
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: etalab-2.0
task_categories:
- text-generation
language:
- fr
size_categories:
- 1K<n<10K
---
This dataset provides a curated collection of preference pairs (chosen vs rejected responses) designed for Direct Preference Optimization (DPO) and related post-training methods.  
It is built from the public [Compar:IA](https://comparia.beta.gouv.fr/) votes dataset, an initiative from the French Ministry of Culture, which gathers real human preferences between model-generated responses.  
The raw conversational data has been transformed into a clean and structured format, making it directly usable for alignment training of large language models. 
To support different training setups and resource constraints, the dataset is released in two versions:

[6k samples](https://huggingface.co/datasets/jpacifico/comparia-dpo-pairs-bt-6k) (this one)  
[13k samples](https://huggingface.co/datasets/jpacifico/comparia-dpo-pairs-bt-13k)

**Key Features** 

- Built from real human feedback (not synthetic data)
- Derived from the french government-backed [votes dataset](https://huggingface.co/datasets/ministere-culture/comparia-votes)
- Enhanced using Bradley-Terry preference modeling
- Ready-to-use DPO format (chosen / rejected pairs)
- Cleaned and filtered for training stability
- Strong focus on French language, with cross-lingual generalization potential

Thanks to [Ministère de la Culture](https://huggingface.co/ministere-culture)

Built by : Jonathan Pacifico, 2026