OmniAlign-V-DPO: 150k Preference Pairs for Multimodal LLM Alignment

OmniAlign-V-DPO datasets contains 150,000 high-quality positive-negative pairs for Direct Preference Optimization (DPO). It is based on the OmniAlign-V datasets and was created by PhoenixZ. The dataset was last updated on March 1, 2025.

Use Cases

• Training multimodal LLMs via Direct Preference Optimization based on the 150k preference pairs.
• Benchmarking model alignment performance using the referenced MM-AlignBench.
• Fine-tuning vision-language models like LLaVANext-OA variants with human preference data.
• Researching methods for enhancing multimodal model alignment.

Strengths

• Contains 150,000 high-quality positive-negative pairs, providing a substantial training resource.
• Is the official dataset from a referenced research paper and GitHub repository.
• Specifically designed for Direct Preference Optimization (DPO), a targeted training method.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known (150k pairs), but specific file formats and data structure details are unknown.
• Freshness should be verified; the last update was March 1, 2025.

Provenance

Source

PhoenixZ

Collection Method

Derived from the OmniAlign-V datasets.

Freshness

Last updated 2025-03-01 09:22:05.