ITG/PlatVR-kto

PlatVR KTO Dataset

This dataset is part of the EVIDENT framework, designed to enhance the creative process of generating background images for virtual reality sets.

Disclaimer

The creation process was done using a crowdsourcing methodology. Therefore, the preferences in the data align with the user group that participated in the process (i.e., these are real preference data).

Dataset Details

This dataset followed a creation process using our fine-tuned model ITG/PlatVR-dpo.

Creation Process

We deployed our fine-tuned model for inference so that multiple users could interact with it and give feedback according to their subjective opinions. We used this frontend web demo to collect real human feedback.

The users interact with the model within a chatbot interface with the following steps for each new chat session:

1. The user sends an initial description to the system.
2. The system returns an enhanced description by adding more details.
3. The user sends instructions for adding, condensing, modifying, rearranging, removing, rephrasing, or overall scene change for the description to the system.
4. The system returns a new description with changes done according to the user input.

The steps (3 and 4) can be repeated as often as the user wants in the same chat. The user can give binary feedback (like/dislike) to the system responses (in steps 2 and 4).

Although it is possible to interact with the system in multiple languages the system always answers in English.

The user sends messages in Spanish in this crowdsourcing process.

We manually filter the data to ensure the adequacy of the samples for our framework.

Example

{
  "id": 1,
  "instruction": "es de noche y hay focos de luz",
  "original_prompt": "A scene of industrial production, featuring towering factories and machinery, with smoke billowing from chimneys and a bustling workforce operating the equipment amidst a backdrop of metal girders and storage facilities. In the distance, groups of workers can be seen conversing and examining blueprints, their focused discussions contributing to the efficient operation of the facility.",
  "modified_prompt": "A scene of industrial production at night, featuring towering factories and machinery, with smoke billowing from chimneys and a bustling workforce operating the equipment amidst a backdrop of metal girders and storage facilities. In the distance, groups of workers can be seen conversing and examining blueprints, their focused discussions contributing to the efficient operation of the facility. The area is illuminated by focused beams of light, highlighting the industrial landscape and casting shadows that add depth to the scene.",
  "label": true
}

Citation

@misc{PlatVR-KTO-Dataset,
  title = {PlatVR-kto: A synthetic KTO dataset about landscape descriptions with crowdsourced preferences},
  author = {ITG},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/ITG/PlatVR-kto}
}