Yangyihui/nav-wam-short-v6-textfix
收藏资源简介:
nav-wam-short v6 (VLM改进文本)是一个消融数据集,专注于文本指令质量。它基于nav-wam-short-v5的收集流程(使用collect_goal_point_segments.py --min-path-complexity-deg 120,无面向障碍物),但进行了以下改进:切换VLM模型从gemini-3.1-flash-lite-preview(已停用)到gemini-2.5-flash(更强);增加了VLM调用的预算以支持完整推理和JSON输出;当VLM验证接受重写时,即使重写删除了数据集标签令牌,也会使用重写(以捕捉InteriorGS的错误标签,例如将发廊的“沙发”实际识别为沙龙椅);当VLM拒绝并返回visible=<名词>时,模板会将数据集标签替换为VLM可见的名词(例如,wardrobe -> cabinet)。该数据集包含214个片段,其中52%的片段获得了VLM改进的文本,相比v5的约30%有所提升。数据集仅包含验证集(val.txt),用于文本质量比较的纯测试集。
nav-wam-short v6 (VLM-enhanced text) is an ablation dataset focused on text instruction quality. It follows the data collection pipeline of nav-wam-short-v5 (using `collect_goal_point_segments.py --min-path-complexity-deg 120`, no obstacle-facing constraints), with the following improvements: 1. Switched the VLM model from gemini-3.1-flash-lite-preview (now discontinued) to gemini-2.5-flash, which offers stronger performance; 2. Increased the VLM inference budget to support full reasoning and JSON output; 3. When the VLM validation accepts a rewritten text, the revised version will be used even if it deletes dataset label tokens, to fix mislabeling issues in InteriorGS (e.g., identifying the "sofa" in a hair salon as a salon chair); 4. When the VLM rejects the original text and returns `visible=<noun>`, the template will replace the dataset label with the noun recognized by the VLM (e.g., wardrobe -> cabinet). This dataset contains 214 segments, 52% of which have VLM-enhanced text, an improvement over the ~30% rate in version v5. The dataset only includes the validation set (val.txt) as a pure test set for text quality comparison.




