File size: 4,473 Bytes
4b3746e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4970332
4b3746e
 
4970332
 
 
 
4b3746e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4970332
 
 
 
 
4b3746e
4970332
 
4b3746e
4970332
4b3746e
 
 
 
 
 
 
 
 
 
 
 
4970332
 
 
4b3746e
 
4970332
 
 
 
 
4b3746e
 
 
 
 
 
 
 
 
 
 
 
 
4970332
 
4b3746e
4970332
4b3746e
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
---
pretty_name: DynamicWAM Training Data
license: apache-2.0
language:
- en
task_categories:
- robotics
size_categories:
- 1M<n<10M
tags:
- robot-manipulation
- imitation-learning
- video-latents
- optical-flow
- safetensors
- domino
- dynamicwam
---

# DynamicWAM training data

This repository contains the packed training corpus for **DynamicWAM (w/o
motion)**, the paper variant previously tracked internally as B0. The public
name is now consistently **DynamicWAM (w/o motion)**.

The corpus is a lossless packed derivative of
[`H-EmbodVis/DOMINO`](https://github.com/H-EmbodVis/DOMINO), not a copy of the
raw RGB dataset. It stores aligned video latents, action/state targets, history
Flow RGB, language embeddings, frame indices, and metadata. The Hub layout
buckets the language shards to satisfy Hugging Face's per-directory file
limit; a supplied materialization tool creates the flat compatibility view
expected by the released training loader without duplicating payload bytes.

## Important naming distinction

“w/o motion” means **without the explicit numeric absolute-motion descriptor
tokens** introduced by the full DynamicWAM model. This dataset still contains
four history optical-flow intervals represented as Flow RGB. Its temporal
contract is `model_policy_index_aligned_v1`; physical simulator timestamps are
not available in this earlier corpus.

For the later exact-simulator-time corpus used by full DynamicWAM, use
[`KhalilGao/DOMINO_absolute_motion_v2`](https://huggingface.co/datasets/KhalilGao/DOMINO_absolute_motion_v2).

## Dataset summary

| Field | Value |
|---|---:|
| Tasks | 35 |
| Episodes | 26,019 |
| Clean episodes | 5,207 |
| Randomized episodes | 20,812 |
| Packed samples | 3,455,662 |
| Main data shards | 346 |
| Payload files | 82,008 |
| Logical payload size | 2,199,796,269,773 bytes (2.00 TiB) |
| Zero-byte files | 0 |
| Dataset fingerprint | `bc20355aa094585499c2ade05d8bcfdb1f082aabc73259ebcded694e3301c425` |

## Repository layout

```text
dynamicwam_wo_motion/
├── dataset.json
├── stats.json
├── action_stats.json
├── episodes.jsonl
├── samples.jsonl
├── shards/
│   ├── shard_*.safetensors
│   └── shard_*.json
└── lang/
    ├── lang.json
    └── shards/
        ├── bucket_000/
        │   ├── shard_000000.safetensors
        │   └── ...
        └── bucket_040/
            └── shard_040654.json
DATASET_RELEASE.json
FILE_LAYOUT.json
PAYLOAD_MANIFEST.jsonl
SHA256SUMS
tools/materialize_training_layout.py
```

This is a custom Safetensors-backed training format, so the Hugging Face
Dataset Viewer is not the authoritative loader. Use the DynamicWAM dataset
loader from the source repository.

## Download

```bash
hf download KhalilGao/DynamicWAM-data \
  --repo-type dataset \
  --local-dir data/DynamicWAM-data

python data/DynamicWAM-data/tools/materialize_training_layout.py \
  data/DynamicWAM-data/dynamicwam_wo_motion
```

The second command creates relative symlinks such as
`lang/shards/shard_000000.safetensors ->
bucket_000/shard_000000.safetensors`. It is idempotent. Use `--mode hardlink`
when symlinks are unsuitable; neither mode duplicates the underlying payload.

The matching checkpoint is
[`KhalilGao/DynamicWAM/checkpoints/dynamicwam_wo_motion/model.pt`](https://huggingface.co/KhalilGao/DynamicWAM/blob/main/checkpoints/dynamicwam_wo_motion/model.pt).

## Integrity and provenance

- Format: `efficient_wam_train_dataset`, version 2.
- Action/state dimension: 14; action chunk size: 16.
- Future latent sequence: 8 frames; composite video size: 384×320.
- History flow: head view, four intervals, 64×64 Farneback flow, P99
  normalization, fixed raw stride.
- The packaged action statistics have SHA-256
  `730a91e41b6a23c70cec49cd222835373ba9364680da5753f4408c5534319b04`,
  exactly matching the released DynamicWAM (w/o motion) checkpoint.
- `PAYLOAD_MANIFEST.jsonl` maps every native logical path to its bucketed Hub
  path and records byte size and source modification time.
- `SHA256SUMS` is generated from the upload-time content hashes and covers the
  complete bucketed payload under `dynamicwam_wo_motion/`.

## License

The packaged derivative is released under Apache-2.0, consistent with the
upstream DOMINO data/code release. Language and video-model-derived features
remain subject to the applicable upstream model terms; no external model
weights are included here.