nusshao/SocialNav-Dataset
SocialNav-SUB数据集制作 Background Raw Dataset地址为 https://huggingface.co/datasets/michaelmunje/SocialNav-SUB 根据原论文,是一个VQA(Video Question Answering)的数据集。 但是作者只提供了Raw Data,没法直接用来训练VLM。此项目为制作openai-chat格式的VQA数据集,用于llama-factory的SFT训练。 前视图提取 作者提供的human-labeled的RGB图像,是front-view和BEV拼接在一起的。 我们运行extract_labeled.py将labeled RGB提取到/labeled_front_view文件夹中,同时保留num of people的信息。 最后,运行downsample.py将图片分辨率降采样为640*360。 问答对构建 运行build_dataset.py即可… See the full description on the dataset page: https://huggingface.co/datasets/nusshao/SocialNav-Dataset.
Update qwenvl_sft.sh
Update sft_configs/lora_sft.yaml
Upload extract_labeled.py
Delete extract_labeled.py
Update dataset_info.json
Update sft_configs/lora_sft.yaml
Upload downsample.py
Upload 660 files
Delete labeled_front_view
Update README.md
Upload 2 files
Delete build_dataset.py
Delete dataset.json
Initial upload of SocialNav dataset
initial commit
