AMALIA-VL-DPO: Multimodal Preference Triplets for Vision-Language Model Alignment

AMALIA-VL-DPO Dataset is a collection of Direct Preference Optimization (DPO) data for the AMALIA-VL project. It contains preference triplets where each row includes a prompt with normalized role-content turns and image placeholders, a chosen assistant response, and a rejected response. The dataset was created by author 'amalia-llm' and was last updated on June 30, 2026.

Use Cases

• Training Direct Preference Optimization (DPO) models based on the described preference triplets.
• Aligning vision-language assistant responses based on human or AI feedback indicated in the data.
• Fine-tuning multimodal language models using the structured prompt-response format with image placeholders.
• Studying preference learning for multimodal conversational agents based on the triplet structure.

Strengths

• Data is structured as preference triplets, which are a standard format for alignment tasks.
• Prompts are normalized with role-content turns and explicit image placeholders for multimodal contexts.
• One subset, 'vl_preference_200k', is noted as being fully derived from a supervised fine-tuning mix.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Row count, column details, and file formats are unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.

Provenance

Source

huggingface

Collection Method

Part of the AMALIA project; data used for AMALIA-VL. Subsets are derived from various sources, including a supervised fine-tuning mix.

Freshness

Last updated 2026-06-30 21:30:51; freshness should be verified.