This dataset contains 12,070 frames collected in the Mangalia harbor Gazebo simulation,
used to fine-tune a vision-language-action (VLA) model for marine navigation.
Built with opencode + DeepSeek V4 Flash — every component
of this project (simulation setup, expert controller, data pipeline, ML training,
inference nodes) was developed using opencode CLI with DeepSeek V4 Flash as the
underlying model.
v2 Changes (Fixed Expert)… See the full description on the dataset page: https://huggingface.co/datasets/MSaalaamaa/marine_vla_dataset_v2.
