OmniAlign-V: 205k Samples for Multimodal LLM Human Preference Alignment

205k high-quality samples for aligning Multimodal Large Language Models with human preferences. The dataset was created by PhoenixZ and is associated with the paper 'OmniAlign-V: Towards Enhanced Alignment of MLLMs with Human Preference'. It was last updated on March 1, 2025.

Use Cases

• Fine-tuning MLLMs for better human preference alignment based on the described high-quality samples.
• Training Direct Preference Optimization (DPO) models using the companion DPO dataset.
• Benchmarking MLLM performance on alignment tasks using the referenced MM-AlignBench.
• Developing new alignment techniques for vision-language models based on the multimodal training data.

Strengths

• Contains 205k high-quality samples.
• Is the official dataset for a published research paper.
• Provides companion resources including a DPO dataset and evaluation benchmarks.

Limitations

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

Provenance

Source

PhoenixZ

Freshness

Last updated 2025-03-01 09:21:45; freshness should be verified.