codemanCheng/physicalword-assets

PhysicalWord last05_mot2_action Assets

This dataset repo contains the private/custom assets needed to train last05_mot2_action on a machine that cannot access the source NAS.

Code repo expected by the migration guide:

• GitHub: https://github.com/zhang-yi-ming/phyword.git
• Branch: mot2_cosmos_ae
• Commit: 1495a9d

Included

This repo intentionally includes:

ckpt/pretrained/LaST0_Pretrain_AE_chunk16/tfmr/
data/libero_training_data_last05_lastest/libero_spatial_20hz_224_dual/
rlbench/

Approximate source sizes:

ckpt/pretrained/LaST0_Pretrain_AE_chunk16/tfmr                         4.0G
data/libero_training_data_last05_lastest/libero_spatial_20hz_224_dual  2.8G
rlbench                                                               25G

Not Included

Cosmos Predict2.5 weights are not included in this repo. Download them from the official gated NVIDIA Hugging Face model after accepting the license:

huggingface-cli download nvidia/Cosmos-Predict2.5-2B \
  --local-dir /mnt/nas/zhangyiming/database/ckpt/pretrained/Cosmos-Predict2.5-2B

The source scripts expect the Cosmos checkpoint at:

/mnt/nas/zhangyiming/database/ckpt/pretrained/Cosmos-Predict2.5-2B/base/pre-trained/d20b7120-df3e-4911-919d-db6e08bad31c_ema_bf16.pt

Already-trained MoT2 evaluation checkpoints are also not included.

Target Layout

Download this dataset repo directly into:

/mnt/nas/zhangyiming/database

Example:

export HF_HUB_ENABLE_HF_TRANSFER=1
mkdir -p /mnt/nas/zhangyiming/database

huggingface-cli download <owner>/physicalword-assets \
  --repo-type dataset \
  --local-dir /mnt/nas/zhangyiming/database

The training JSON files contain absolute paths under /mnt/nas/zhangyiming/database. If the assets are downloaded elsewhere, create a symlink:

mkdir -p /mnt/nas/zhangyiming
ln -sfn /actual/database/path /mnt/nas/zhangyiming/database

Asset Notes

Action Expert

Expected target path:

/mnt/nas/zhangyiming/database/ckpt/pretrained/LaST0_Pretrain_AE_chunk16/tfmr

Files:

config.json
model.safetensors
preprocessor_config.json
processor_config.json
special_tokens_map.json
tokenizer.json
tokenizer_config.json

This directory is used as both the action expert and processor/tokenizer source by the MoT2 scripts. Do not replace it with Janus-Pro-1B.

RLBench

Expected target root:

/mnt/nas/zhangyiming/database/rlbench

The full source rlbench directory is included. The key training files used by the current scripts include:

rlbench/train/json/train_action_chunk1_sumpos_lastrot.json
rlbench/train/json/train_action_chunk1_sumpos_lastrot_statistics.json
rlbench/train/json/cosmos_text_cache_rlbench_keyframe/
rlbench/train/images/

The JSON references image paths such as:

/mnt/nas/zhangyiming/database/rlbench/train/images/...

LIBERO

Expected target path:

/mnt/nas/zhangyiming/database/data/libero_training_data_last05_lastest/libero_spatial_20hz_224_dual

Included files/directories:

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/

The JSON references video paths such as:

/mnt/nas/zhangyiming/database/data/libero_training_data_last05_lastest/libero_spatial_20hz_224_dual/videos/...

Main Training Entrypoints

RLBench:

cd /mnt/nas/zhangyiming/last05_beta/last05_mot2_action
bash scripts/train_mot2_rlbench_keyframe.sh

LIBERO:

cd /mnt/nas/zhangyiming/last05_beta/last05_mot2_action
bash scripts/train_mot2.sh