---
pretty_name: Doom Frame Dataset
tags:
- doom
- vizdoom
- reinforcement-learning
- imitation-learning
- webdataset
configs:
- config_name: preview
  data_files:
  - split: train
    path: data/train-000000.tar
- config_name: full
  data_files:
  - split: train
    path: data/train-*.tar
---

# DoomFrameDataset

DoomFrameDataset is a ViZDoom frame-action dataset generated from policy rollouts. It is packaged as WebDataset tar shards for streaming training, imitation learning, behavior cloning, and offline reinforcement-learning experiments.

The dataset contains RGB game frames paired with the action selected by the rollout policy and per-step metadata such as reward, episode id, step id, terminal flag, and value estimate.

## Dataset Size

| Config | Files | Samples | Intended use |
| --- | ---: | ---: | --- |
| `preview` | 1 shard | ~79k | Hugging Face preview and quick sanity checks |
| `full` | 31 shards | 2,398,745 | Training and full streaming reads |

The packaged dataset is about 68 GB.

## Files

```text
data/
  train-000000.tar
  train-000001.tar
  ...
  train-000030.tar
action_map.json
README.md
```

Each tar shard contains paired files with the same numeric key:

```text
000000000000.png
000000000000.json
000000000001.png
000000000001.json
...
```

The PNG is the game frame. The JSON is the metadata for that frame.

## Sample Metadata

```json
{
  "action_id": 1,
  "action_name": "TURN_RIGHT",
  "action_vector": [0.0, 0.0, 0.0, 0.0, 1.0, 0.0],
  "curriculum_level": 1,
  "done": false,
  "episode": 1,
  "frame_path": "frames/episode_001/step_000000.png",
  "global_step": 0,
  "reward": 0.0,
  "source_frame_path": "frames/episode_001/step_000000.png",
  "step": 0,
  "value": 1.7968196868896484,
  "webdataset_key": "000000000000"
}
```

See `action_map.json` for the full action id, action name, and action vector mapping.

## Load The Preview Config

Use `preview` when you only want to verify the dataset or inspect examples in the Hugging Face Dataset Viewer.

```python
from datasets import load_dataset

ds = load_dataset(
    "brahmandam/DoomFrameDataset",
    "preview",
    split="train",
    streaming=True,
)

sample = next(iter(ds))
print(sample.keys())
print(sample["json"])
image = sample["png"]
```

## Stream The Full Dataset

Use `full` for training.

```python
from datasets import load_dataset

ds = load_dataset(
    "brahmandam/DoomFrameDataset",
    "full",
    split="train",
    streaming=True,
)

for sample in ds:
    image = sample["png"]
    metadata = sample["json"]
    action_id = metadata["action_id"]
    break
```

You can also read the shards directly with WebDataset:

```python
import webdataset as wds

urls = "https://huggingface.co/datasets/brahmandam/DoomFrameDataset/resolve/main/data/train-{000000..000030}.tar"

dataset = (
    wds.WebDataset(urls)
    .decode("pil")
    .to_tuple("png", "json")
)

image, metadata = next(iter(dataset))
```

## Notes

The `preview` config intentionally points to a single shard so the Hub can inspect a small part of the dataset without processing the full 68 GB. For training, use the `full` config.

This dataset was generated from automated ViZDoom policy rollouts. It should be treated as gameplay observation/action data, not human demonstrations.