SauerkrautLM-Doom-MultiVec-31k

31,645 human gameplay demonstration frames for training the 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

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

# 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, 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

@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.