MAmmoTH-VL-Instruct-12M: Multimodal Instruction Examples for Vision-Language Pre-training

MAmmoTH-VL-Instruct-12M is a dataset of interleaved multimodal examples adapted for the modality-aware continual pre-training of MoCa models. The dataset, created by moca-embed, was last updated on July 1, 2025. It is structured for visual question answering (VQA) tasks by concatenating prompts and responses.

Use Cases

• Pre-training vision-language models based on interleaved multimodal examples mentioned in the description
• Instruction tuning for visual question answering based on the VQA-style structure
• Benchmarking model performance on modality-aware tasks based on the dataset's adaptation for MoCa models

Strengths

• Dataset is explicitly designed for modality-aware continual pre-training of MoCa models
• Last updated on 2025-07-01, indicating recent maintenance

Limitations

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

Provenance

Source

huggingface, author moca-embed

Collection Method

Adapted from MAmmoTH-VL-Instruct-12M by concatenating prompts and responses for pre-training

Freshness

Last updated 2025-07-01 03:03:59