gabrielsunhyuck/IALA_B

IALA_B日间单标JSON至JSON数据集包含1000个样本，输入是VLM观测JSON（包括body/topmark属性，可能包含同义词/未知），输出是LLM标准化JSON（包括mark_type, safe_direction, hazards, consistency, finalized_attributes等）。数据集假设使用了IALA-B标准（红色代表右舷，绿色代表左舷）。共有1500个样本，分为9个类别，包括东向标、孤立危险、左舷、右舷、北向标、安全水域、南向标、特殊标和西向标。输入数据通过同义词进行增强，输出被标准化到一个固定的词汇表；冲突/未知的输入在finalized_attributes中被降级为unknown并在consistency.flagged_inconsistent中记录。

The IALA_B Daytime, Single-Mark JSON→JSON Dataset contains 1,000 samples with inputs as VLM observation JSON (including body/topmark attributes, may include synonyms/unknown) and outputs as LLM normalized JSON (including mark_type, safe_direction, hazards, consistency, finalized_attributes, etc.). The dataset assumes the use of IALA-B standards (red for starboard, green for port). There are a total of 1,500 samples divided into 9 classes: cardinal_east, isolated_danger, lateral_port, lateral_starboard, cardinal_north, safe_water, cardinal_south, special_mark, and cardinal_west. Input data is augmented with synonyms, and outputs are normalized to a fixed vocabulary; conflicting/unknown inputs are downgraded to unknown in finalized_attributes and recorded in consistency.flagged_inconsistent.