physicalword-assets
收藏资源简介:
本数据集(PhysicalWord last05_mot2_action Assets)是一个用于训练last05_mot2_action模型的专用资产集合,旨在为无法直接访问原始网络存储(NAS)的环境提供必要的训练资源。它并非原始标注数据集,而是一个包含预训练模型、处理后的训练数据及配套文件的资产包。核心资产包括动作专家模型(位于ckpt/pretrained/目录,作为动作先验知识源)、RLBench训练数据(完整的rlbench目录,关联机器人操作任务)和LIBERO训练数据(位于data/libero_training_data_last05_lastest/libero_spatial_20hz_224_dual目录,关联具身AI与视觉语言任务)。数据规模约31.8GB,涵盖模型检查点文件、JSON元数据、图像帧和视频文件。主要用途是作为运行特定训练脚本的前提条件,用于训练与多模态动作理解和生成相关的模型。注意:不包含Cosmos Predict2.5-2B模型权重,需从官方渠道下载;部署对绝对路径有强依赖,需创建符号链接以确保资源正确定位。
This dataset repository (PhysicalWord last05_mot2_action Assets) is a specialized asset collection for training the last05_mot2_action model, designed to provide necessary training resources for environments without direct access to the original network storage (NAS). It is not an original annotated dataset but an asset package containing pre-trained models, processed training data, and supporting files. Core assets include: 1) Action Expert model (located in the ckpt/pretrained/ directory, serving as a source of action prior knowledge), 2) RLBench training data (complete rlbench directory, associated with robot manipulation tasks), and 3) LIBERO training data (located in data/libero_training_data_last05_lastest/libero_spatial_20hz_224_dual directory, associated with embodied AI and vision-language tasks). The total asset size is approximately 31.8GB, covering model checkpoint files, structured JSON metadata, image frames, and video files. The primary use is as a prerequisite for running specific training scripts (e.g., scripts/train_mot2_rlbench_keyframe.sh and scripts/train_mot2.sh) to train models related to multimodal action understanding and generation. Important notes: The repository does not include weights for the Cosmos Predict2.5-2B model, which must be downloaded separately from official sources after accepting the license. Asset deployment strongly depends on absolute paths (/mnt/nas/zhangyiming/database); if the download location differs, symbolic links must be created as required to ensure training scripts correctly locate resources.
数据集概述
该数据集为 last05_mot2_action 模型的训练提供了必要的私有/定制资产。数据集地址为:https://huggingface.co/datasets/codemanCheng/physicalword-assets
包含内容
数据集包含以下三个主要部分:
-
预训练模型检查点:
- 路径:
ckpt/pretrained/LaST0_Pretrain_AE_chunk16/tfmr - 大小:约 4.0 GB
- 用途:同时用作动作专家(Action Expert)以及分词器(Processor/Tokenizer)的来源。
- 文件列表:
config.json、model.safetensors、preprocessor_config.json、processor_config.json、special_tokens_map.json、tokenizer.json、tokenizer_config.json。
- 路径:
-
LIBERO 训练数据:
- 路径:
data/libero_training_data_last05_lastest/libero_spatial_20hz_224_dual - 大小:约 2.8 GB
- 包含文件/目录:
train_with_atomic_action.json、train.json、train_rewritten_prompts.json、train_statistics.json、videos/、cosmos_text_cache/、cosmos_text_cache_raw_full_concat/、cosmos_text_cache_rewritten_prompts_raw_full_concat/。 - 说明:JSON 文件中引用的视频路径位于
/mnt/nas/zhangyiming/database/data/libero_training_data_last05_lastest/libero_spatial_20hz_224_dual/videos/...。
- 路径:
-
RLBench 训练数据:
- 路径:
rlbench/ - 大小:约 25 GB
- 包含关键训练文件:
rlbench/train/json/train_action_chunk1_sumpos_lastrot.jsonrlbench/train/json/train_action_chunk1_sumpos_lastrot_statistics.jsonrlbench/train/json/cosmos_text_cache_rlbench_keyframe/rlbench/train/images/
- 说明:JSON 文件中引用的图片路径位于
/mnt/nas/zhangyiming/database/rlbench/train/images/...。
- 路径:
不包含内容
- Cosmos Predict2.5 权重:需要从官方 Hugging Face 模型库 (
nvidia/Cosmos-Predict2.5-2B) 下载,并接受许可协议。 - MoT2 评估检查点:已训练好的评估模型检查点未包含在内。
使用说明
- 目标布局:建议将本数据集直接下载至
/mnt/nas/zhangyiming/database目录。 - 路径兼容:训练 JSON 文件中的路径均为绝对路径,基于
/mnt/nas/zhangyiming/database。如果下载到其他位置,需要创建符号链接。 - 关联代码库:
- GitHub:
https://github.com/zhang-yi-ming/phyword.git - 分支:
mot2_cosmos_ae - 提交:
1495a9d
- GitHub:
主要训练入口
- RLBench:在
/mnt/nas/zhangyiming/last05_beta/last05_mot2_action目录下执行bash scripts/train_mot2_rlbench_keyframe.sh。 - LIBERO:在
/mnt/nas/zhangyiming/last05_beta/last05_mot2_action目录下执行bash scripts/train_mot2.sh。




