README.md

metadata

hackathon: Build Small (2026)
title: Dreadzone
emoji: 💬
colorFrom: yellow
colorTo: red
sdk: gradio
sdk_version: 6.5.1
app_file: app.py
pinned: false
suggested_hardware: t4-small
license: artistic-2.0
short_description: Backrooms-inspired local GGUF experience
team:
  - grimjim
tags:
  - track:wood
  - sponsor:openai
  - sponsor:nvidia
  - achievement:offgrid
  - achievement:llama
social_media_post: >-
  https://www.linkedin.com/posts/jim-lai-038249_i-participated-in-the-build-small-hackathon-share-7472113354073853952-LA39/

An entry for the Build Small Hackathon (2026) The track taken: Thousand Token Wood

Dreadzone is a Backrooms-inspired interactive fiction prototype that runs a local GGUF model with llama-cpp-python and Gradio ChatInterface.

The app downloads unsloth/NVIDIA-Nemotron-3-Nano-4B-GGUF automatically on first launch and streams responses from NVIDIA-Nemotron-3-Nano-4B-Q5_K_M.gguf.

No hosted inference API, OAuth token, secrets, or external inference services are used. The default dependency pin uses the CUDA 12.4 llama-cpp-python wheel for GPU Spaces.

The Python app owns the lightweight game state: coordinates, turn count, sanity, zone profile, and encounter rolls. The model receives hidden state each turn and narrates the result without exposing coordinates or mechanics. There are a few surprises to keep players on their toes.

Runtime settings

The defaults are intentionally conservative while enabling GPU offload:

• N_CTX=2048
• N_BATCH=128
• MAX_HISTORY_TURNS=6
• GAME_SEED=dreadzone
• N_THREADS defaults to one fewer than the detected CPU count
• N_GPU_LAYERS=-1 offloads all possible layers to GPU
• ENABLE_THINKING=false renders the model chat template with thinking disabled

You can override the model or runtime settings with Space variables:

• MODEL_REPO
• MODEL_FILE
• MODEL_DIR
• GAME_SEED
• N_CTX
• N_BATCH
• N_THREADS
• N_GPU_LAYERS
• ENABLE_THINKING
• MAX_HISTORY_TURNS

Author

grimjim@huggingface

Assisted by Codex