VAGOsolutions/SauerkrautLM-Doom-MultiVec-31k

---
dataset_info:
features:
- name: text
dtype: string
- name: depth_bins
sequence: int64
- name: scores
sequence: float64
- name: input_ids
sequence: int64
- name: attention_mask
sequence: int64
- name: depth_ids
sequence: int64
splits:
- name: train
num_examples: 31645
license: apache-2.0
task_categories:
- text-classification
size_categories:
- 10K<n<100K
language:
- en
tags:
- doom
- game-ai
- ascii
- vizdoom
- human-demonstrations
- depth-data
- SauerkrautLM
pretty_name: SauerkrautLM Doom MultiVec 31K
---

<img src="Logo.png" width="500" height="auto">

# SauerkrautLM-Doom-MultiVec-31k

**31,645 human gameplay demonstration frames for training the [SauerkrautLM-Doom-MultiVec-1.3M](https://huggingface.co/VAGOsolutions/SauerkrautLM-Doom-MultiVec-1.3M) DOOM action classifier.**

This dataset was recorded by a human player in VizDoom's SPECTATOR mode across 4 recording sessions totaling approximately 2 hours of gameplay in the `defend_the_center` scenario. Each frame includes the ASCII game view, real VizDoom depth buffer data, and soft action labels derived from keyboard input.

---

## Dataset Structure

Each sample contains:

| Field | Type | Description |
|-------|------|-------------|
| `text` | string | 40x25 ASCII frame (~1024 characters), brightness-encoded |
| `depth_bins` | list[int] | VizDoom depth buffer quantized to 16 bins per token position |
| `scores` | list[float] | 4-dim soft action scores: [shoot, move_forward, turn_left, turn_right] |
| `input_ids` | list[int] | Pre-tokenized with 75-token character-level vocabulary |
| `attention_mask` | list[int] | Attention mask aligned to input_ids |
| `depth_ids` | list[int] | Depth bin IDs aligned to token positions (16 = no depth / padding) |

### Soft Action Scores

Action labels are **soft distributions**, not hard one-hot labels. When the human presses multiple keys simultaneously (e.g., forward + shoot), both actions receive high scores (0.85), while inactive actions receive a baseline of 0.05. This provides richer supervision for KL-divergence training.

### ASCII Encoding

Each frame uses 10 brightness characters: `" .:-=+*#%@"` (dark to bright). Bright characters indicate nearby solid objects; dark characters indicate distant or empty areas. Row separators (`\n`) preserve the 2D spatial layout.

---

## Recording Setup

| Setting | Value |
|---------|-------|
| **Scenario** | `defend_the_center` (circular arena, enemies from all directions) |
| **Resolution** | 640x480 with HUD enabled |
| **Frame skip** | 4 (one sample per 4 game tics, ~114ms real-time) |
| **Controls** | Native DOOM keyboard (arrow keys + Ctrl) |
| **Actions** | 4 discrete: shoot, move_forward, turn_left, turn_right |
| **Depth source** | VizDoom depth buffer, quantized to 16 bins |
| **Recording sessions** | 4 sessions, 80+ episodes, ~2 hours total |
| **Total frames** | 31,645 |

---

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("VAGOsolutions/SauerkrautLM-Doom-MultiVec-31k")
train = dataset["train"]

print(f"Samples: {len(train)}")
print(f"Features: {list(train.features.keys())}")

# Inspect a sample
sample = train[0]
print(f"ASCII frame length: {len(sample['text'])} chars")
print(f"Action scores: {sample['scores']}")
print(f"Depth bins (first 10): {sample['depth_bins'][:10]}")
```

### Train with this dataset

```bash
# Clone the project
git clone https://github.com/VAGOsolutions/doom-multivec.git
cd doom-multivec
pip install -e ".[dev]"

# Train the classifier
python scripts/train_classifier.py \
--data VAGOsolutions/SauerkrautLM-Doom-MultiVec-31k \
--output output/my-model \
--epochs 10 \
--batch-size 32 \
--lr 3e-4
```

---

## Associated Model

This dataset was used to train **[SauerkrautLM-Doom-MultiVec-1.3M](https://huggingface.co/VAGOsolutions/SauerkrautLM-Doom-MultiVec-1.3M)**, a 1.3M parameter ModernBERT-Hash classifier that achieves 178 frags in 10 episodes of VizDoom's `defend_the_center`, outperforming GPT-4o-mini, Nemotron-120B, Qwen3.5-27B, and Gemini Flash Lite combined.

---

## Citation

```bibtex
@misc{SauerkrautLM-Doom-MultiVec,
title={SauerkrautLM-Doom-MultiVec-1.3M: Playing DOOM with 1.3M Parameters},
author={David Golchinfar and Daryoush Vaziri and Alexander Marquardt},
url={https://huggingface.co/VAGOsolutions/SauerkrautLM-Doom-MultiVec-1.3M},
year={2026}
}
```

---

## License

Apache 2.0 License.

DOOM is a registered trademark of id Software LLC. This project is not affiliated with or endorsed by id Software.