README.md

---
license: other
---

# multi3d_games Latent Dataset

Partial Multi3D games latent dataset uploaded to
`mignonjia/multi3d_games`.

This dataset uses the expanded parquet layout rather than tar archive shards.
It contains latent/video-conditioning rows for four gameplay videos from
Horizon Forbidden West, Dark Souls Remastered, and Code Vein.

## Layout

```text
README.md
map_style_cache/file_info.pkl
action_latent/node1/manifest.jsonl
action_latent/node1/manifest.rankNNN.jsonl
action_latent/node1/dist_merged.ok
action_latent/node1/dist_done/rankNNN.done
action_latent/node1/combined_parquet_dataset/rankNNN/worker_0/*.parquet
```

The uploaded repo contains one node (`node1`) split across 8 ranks:
`rank000` through `rank007`.

## Size and Counts

- Total uploaded data size: about 224.7 GB
- Uploaded files: 1,715 data files plus `README.md` and `.gitattributes`
- Parquet files: 1,696
- Parquet rows / samples: 13,551
- Manifest rows:
  - `manifest.jsonl`: 13,551 rows
  - rank manifests: 13,551 rows total
- Rank parquet layout:
  - `rank000`: 212 parquet files, 1,694 rows, 26.17 GiB
  - `rank001`: 212 parquet files, 1,694 rows, 26.13 GiB
  - `rank002`: 212 parquet files, 1,694 rows, 26.14 GiB
  - `rank003`: 212 parquet files, 1,694 rows, 26.16 GiB
  - `rank004`: 212 parquet files, 1,694 rows, 26.15 GiB
  - `rank005`: 212 parquet files, 1,694 rows, 26.17 GiB
  - `rank006`: 212 parquet files, 1,694 rows, 26.16 GiB
  - `rank007`: 212 parquet files, 1,693 rows, 26.16 GiB

## Source Videos

The rows come from four source videos:

| idx | video_id | game | rows | shard |
| --- | --- | --- | ---: | --- |
| 2246 | `KchWtQyuyvU` | Horizon Forbidden West | 3,396 | `SHARD_0006` |
| 8224 | `d_lTaTapecI` | Dark Souls Remastered | 3,969 | `SHARD_0026` |
| 12601 | `m2Nt3DVfYqk` | Code Vein | 1,847 | `SHARD_0040` |
| 18529 | `soS-p5-Oh7A` | Dark Souls Remastered | 4,339 | `SHARD_0059` |

## Parquet Schema

Each parquet row stores byte arrays plus explicit shape and dtype metadata.
The main fields are:

- `id`: sample id, matching manifest ids
- `vae_latent_bytes`, `vae_latent_shape`, `vae_latent_dtype`
- `clip_feature_bytes`, `clip_feature_shape`, `clip_feature_dtype`
- `first_frame_latent_bytes`, `first_frame_latent_shape`,
  `first_frame_latent_dtype`
- `mouse_cond_bytes`, `mouse_cond_shape`, `mouse_cond_dtype`
- `keyboard_cond_bytes`, `keyboard_cond_shape`, `keyboard_cond_dtype`
- `pil_image_bytes`, `pil_image_shape`, `pil_image_dtype`
- `file_name`, `caption`, `media_type`, `width`, `height`, `num_frames`,
  `duration_sec`, `fps`

Example row metadata:

- `vae_latent_shape`: `[16, 21, 60, 104]`, dtype `float32`
- `first_frame_latent_shape`: `[16, 21, 60, 104]`, dtype `float32`
- `clip_feature_shape`: `[257, 1280]`, dtype `float32`
- `mouse_cond_shape`: `[81, 2]`, dtype `float32`
- `keyboard_cond_shape`: `[81, 6]`, dtype `float32`
- `media_type`: `video`
- `width`: 480
- `height`: 832
- `num_frames`: 21
- `duration_sec`: 2.7
- `fps`: 30.0

## Processing Notes

- Source local root before upload:
  `/mnt/weka/home/hao.zhang/alex/wm-lab/datas/datasets/multi3d-partial`
- The dataset was uploaded directly with `hf upload-large-folder`, preserving
  the expanded parquet paths.
- The original local upload command used 8 workers and committed all
  1,715 files successfully.
- Multi3D mouse up/down convention was corrected before upload by flipping
  `mouse_cond[:, 0]`.
- The mouse flip was validated over all 1,696 parquet files and 13,551 rows.
  The final scan showed the expected swapped axis-0 sign counts relative to the
  pre-flip baseline.

## Download

Download the full dataset:

```bash
hf download mignonjia/multi3d_games --repo-type dataset --local-dir multi3d_games
```

Download one rank only:

```bash
hf download mignonjia/multi3d_games \
  --repo-type dataset \
  --include 'action_latent/node1/combined_parquet_dataset/rank000/**' \
  --local-dir multi3d_games_rank000
```

## Reading Arrays

The array fields are stored as raw bytes. Reconstruct them using the matching
`*_shape` and `*_dtype` columns:

```python
import numpy as np
import pandas as pd

df = pd.read_parquet("action_latent/node1/combined_parquet_dataset/rank000/worker_0/data_chunk_0.parquet")
row = df.iloc[0]

vae = np.frombuffer(row["vae_latent_bytes"], dtype=np.dtype(row["vae_latent_dtype"]))
vae = vae.reshape(tuple(row["vae_latent_shape"]))

mouse = np.frombuffer(row["mouse_cond_bytes"], dtype=np.dtype(row["mouse_cond_dtype"]))
mouse = mouse.reshape(tuple(row["mouse_cond_shape"]))

keyboard = np.frombuffer(row["keyboard_cond_bytes"], dtype=np.dtype(row["keyboard_cond_dtype"]))
keyboard = keyboard.reshape(tuple(row["keyboard_cond_shape"]))
```

## Verification

- Hugging Face repo after upload contained 1,716 files:
  `.gitattributes` plus 1,715 uploaded dataset files.
- Upload log final state:
  - hashed: 1,715 / 1,715
  - pre-uploaded: 1,697 / 1,697
  - committed: 1,715 / 1,715
  - committed bytes: 224.7 GB / 224.7 GB