README.md

metadata

title: ID
emoji: 🔍
colorFrom: gray
colorTo: blue
sdk: gradio
sdk_version: 6.18.0
python_version: '3.12'
app_file: app.py
pinned: false
tags:
  - track:wood
  - sponsor:openbmb
  - sponsor:openai
  - achievement:offgrid
  - achievement:offbrand
  - achievement:llama

ID

A small Gradio build of the LLM-driven investigation game. Question suspects, search rooms, catch contradictions, and make an accusation.

Small-model, no-API by default

The Space runs entirely on a small local model — OpenBMB MiniCPM4.1-8B (~8B params) through the llama.cpp runtime. No cloud API and no user credentials are needed; the GGUF is downloaded from the Hub on first launch.

Recommended hardware: ZeroGPU (set it in the Space's Settings → Hardware). Generation runs inside an @spaces.GPU function, so it uses the GPU on ZeroGPU and falls back to CPU/Metal elsewhere. Worlds are pre-generated and committed under worlds/, so play time only runs the light tiers (character, extractor, guard, environment, judge).

Optional configuration (Space variables)

• F_ID_GGUF_REPO / F_ID_GGUF_FILE: override the model repo / GGUF glob (defaults to openbmb/MiniCPM4.1-8B-GGUF, a Q4_K_M quant)
• F_ID_N_GPU_LAYERS: layers to offload (-1 = all, 0 = CPU only)
• F_ID_CTX: context window (default 16384)
• F_ID_ENABLE_THINKING: enable the model's hybrid <think> reasoning (default off; reasoning is disabled to keep turns fast)

Using a hosted API instead (optional)

The engine is provider-agnostic. To point at an OpenAI-compatible endpoint set:

• F_ID_PROVIDER: openai, openrouter, or local (default local)
• F_ID_MODEL: model name
• F_ID_BASE_URL: OpenAI-compatible API base URL
• F_ID_API_KEY_ENV: name of the Space secret holding the key

More

A CLI version is also available, you can find it here together with the full repository: GitHub - f-id

Demo video link: https://youtu.be/bjLddFEGX9I Social post link: https://www.reddit.com/r/huggingface/comments/1u5l1dp/fid_an_llmdriven_investigation_game/