CoolFace
Datasetpublic

1x-technologies/world_model_tokenized_data

1X World Model Compression Challenge Dataset This repository hosts the dataset for the 1X World Model Compression Challenge. huggingface-cli download 1x-technologies/worldmodel --repo-type dataset --local-dir data Updates Since v1.1 Train/Val v2.0 (~100 hours), replacing v1.1 Test v2.0 dataset for the Compression Challenge Faces blurred for privacy New raw video dataset (CC-BY-NC-SA 4.0) at worldmodel_raw_data Example scripts now split into:… See the full description on the dataset page: https://huggingface.co/datasets/1x-technologies/world_model_tokenized_data.

sourceHugging Faceapache-2.0updated 1y agoView on Hugging Face
34likes2.7kdownloads
Dataset Card

1X World Model Compression Challenge Dataset

This repository hosts the dataset for the 1X World Model Compression Challenge.

bash
huggingface-cli download 1x-technologies/worldmodel --repo-type dataset --local-dir data

Updates Since v1.1

  • —Train/Val v2.0 (~100 hours), replacing v1.1
  • —Test v2.0 dataset for the Compression Challenge
  • —Faces blurred for privacy
  • —New raw video dataset (CC-BY-NC-SA 4.0) at worldmodel_raw_data
  • —Example scripts now split into:
  • —cosmos_video_decoder.py — for decoding Cosmos Tokenized bins
  • —unpack_data_test.py — for reading the new test set
  • —unpack_data_train_val.py — for reading the train/val sets

Train & Val v2.0

Format

Each split is sharded:

  • —video_{shard}.bin — NVIDIA Cosmos Tokenizer discrete DV8×8×8 tokens at 30 Hz
  • —segment_idx_{shard}.bin — segment boundaries
  • —states_{shard}.bin — np.float32 states (see below)
  • —metadata.json / metadata_{shard}.json — overall vs. per‐shard metadata

Test v2.0

We provide a 450 sample test_v2.0 dataset for the World Model Compression Challenge with a similar structure (video_{shard}.bin, states_{shard}.bin). Use:

  • —unpack_data_test.py to read the test set
  • —unpack_data_train_val.py to read train/val ---

State Index Definition (New)

 0: HIP_YAW
 1: HIP_ROLL
 2: HIP_PITCH
 3: KNEE_PITCH
 4: ANKLE_ROLL
 5: ANKLE_PITCH
 6: LEFT_SHOULDER_PITCH
 7: LEFT_SHOULDER_ROLL
 8: LEFT_SHOULDER_YAW
 9: LEFT_ELBOW_PITCH
10: LEFT_ELBOW_YAW
11: LEFT_WRIST_PITCH
12: LEFT_WRIST_ROLL
13: RIGHT_SHOULDER_PITCH
14: RIGHT_SHOULDER_ROLL
15: RIGHT_SHOULDER_YAW
16: RIGHT_ELBOW_PITCH
17: RIGHT_ELBOW_YAW
18: RIGHT_WRIST_PITCH
19: RIGHT_WRIST_ROLL
20: NECK_PITCH
21: Left hand closure (0= open, 1= closed)
22: Right hand closure (0= open, 1= closed)
23: Linear Velocity
24: Angular Velocity

Previous v1.1

  • —video.bin — 16×16 patches at 30Hz, quantized
  • —segment_ids.bin — segment boundaries
  • —actions/ folder storing multiple .bins for states, closures, etc.

v1.1 Joint Index

   {
        0: HIP_YAW
        1: HIP_ROLL
        2: HIP_PITCH
        3: KNEE_PITCH
        4: ANKLE_ROLL
        5: ANKLE_PITCH
        6: LEFT_SHOULDER_PITCH
        7: LEFT_SHOULDER_ROLL
        8: LEFT_SHOULDER_YAW
        9: LEFT_ELBOW_PITCH
        10: LEFT_ELBOW_YAW
        11: LEFT_WRIST_PITCH
        12: LEFT_WRIST_ROLL
        13: RIGHT_SHOULDER_PITCH
        14: RIGHT_SHOULDER_ROLL
        15: RIGHT_SHOULDER_YAW
        16: RIGHT_ELBOW_PITCH
        17: RIGHT_ELBOW_YAW
        18: RIGHT_WRIST_PITCH
        19: RIGHT_WRIST_ROLL
        20: NECK_PITCH
    }
  
A separate `val_v1.1` set is available.

---

## Provided Checkpoints

- `magvit2.ckpt` from [MAGVIT2](https://github.com/TencentARC/Open-MAGVIT2) used in v1.1
- For v2.0, see [NVIDIA Cosmos Tokenizer](https://github.com/NVIDIA/Cosmos-Tokenizer); we supply `cosmos_video_decoder.py`.

---

## Directory Structure Example

trainv1.1/ valv1.1/ trainv2.0/ valv2.0/ testv2.0/ ├── video{shard}.bin ├── states{shard}.bin ├── ... ├── metadata{shard}.json cosmosvideodecoder.py unpackdatatest.py unpackdatatrain_val.py


**License**: [Apache-2.0](./LICENSE)  
**Author**: 1X Technologies
1x-technologies/world_model_tokenized_data · CoolFace