Spaces:
Running
Running
| title: Case Zero | |
| emoji: 🕵️ | |
| colorFrom: indigo | |
| colorTo: yellow | |
| sdk: docker | |
| app_port: 7860 | |
| pinned: true | |
| license: apache-2.0 | |
| models: | |
| - Qwen/Qwen2.5-1.5B-Instruct | |
| tags: | |
| - build-small-hackathon | |
| - thousand-token-wood | |
| - off-the-grid | |
| - llama-champion | |
| - off-brand | |
| - tiny-titan | |
| - field-notes | |
| - sharing-is-caring | |
| - llama-cpp | |
| - detective-game | |
| - text-generation | |
| - tts | |
| short_description: Gather evidence and solve cases in this pixel noir ! | |
| # 🕵️ Case Zero — the AI *is* the detective game | |
| **A brand-new mystery, written and acted by a 1.5B model, every single time.** | |
| No scripted cases. No content library. A single small local model invents the whole | |
| thing — a homicide, a heist, a fraud, a blackmail ring, an arson, a disappearance — the | |
| victim, the suspects, their secrets and motives, the timeline, the evidence, and the one | |
| who did it — then **role-plays every suspect live**. They remember what you asked. They | |
| lie to your face. And when you slap down the right piece of evidence, you watch the lie | |
| **crack in real time**. | |
| > Interrogate. Investigate. Accuse. One of them is guilty. Prove it. | |
| ## ✨ The moment that sells it | |
| Search the rooms, find a clue that contradicts a suspect's alibi, **present it**, and | |
| their story falls apart on screen — stress spikes, the alibi breaks, the truth leaks. | |
| Then name the killer, cite your proof, and get a scored verdict with a "Director's Cut" | |
| walkthrough of how the crime really went down. | |
| ## 🧠 How it works | |
| | Layer | What it does | | |
| |---|---| | |
| | **Model** — Qwen2.5-1.5B-Instruct (GGUF) | The whole game. Runs in-process on the CPU through **llama.cpp** (`llama-cpp-python`) — no server, no GPU, no remote endpoint. | | |
| | **Generation** | The model authors every case as JSON; deterministic Python only wires the *structure* (who's guilty, who was where) so the mystery is always solvable. | | |
| | **Solver** | A fairness referee: single culprit, a breakable alibi, every innocent cleared, and a discoverability gate so the key clue is always findable in play. | | |
| | **Director** | Whether a lie gets caught is decided by **ground truth, not the model** — so the win condition is immune to prose (a jailbroken "just tell me who did it" earns nothing). | | |
| | **Voice** — Supertonic | Each suspect gets a distinct, gender-matched on-device voice, synthesized **sentence-by-sentence as the reply streams**. | | |
| | **Art** | Procedural pixel-art portraits, rooms, and evidence — rendered **client-side on canvas** at one integer-scaled density (so the server spends ~0 CPU on visuals). | | |
| | **UI** | A custom **pixel-art noir SPA (Preact)**, 12 screens, served **100% through `gradio.Server`** (Gradio 6 "Server mode") — the built bundle as static files plus the JSON/SSE `/api` routes, all in one process. No separate frontend host. | | |
| The model does all the creative work. Deterministic code is only guardrails and a | |
| reliability layer — it never writes story, character, or dialogue. | |
| ## 🏆 Built for the Build Small Hackathon | |
| **Track: Thousand Token Wood** — a game that is genuinely impossible without AI: the | |
| model doesn't power a feature, it *authors and performs the entire mystery*. | |
| - **Off the Grid:** zero cloud APIs. The LLM and the TTS run **in-process on the Space's | |
| CPU**; a network-guarded test plays a full case with all non-loopback sockets blocked. | |
| - **Llama Champion:** the model runs through the **llama.cpp** runtime, in-process — no | |
| server, no remote endpoint. | |
| - **Off-Brand:** a fully custom 12-screen pixel-art noir SPA served through | |
| **`gr.Server`** — procedural scenes, exhibit art, and sprites, all canvas-drawn. | |
| - **Tiny Titan (≤4B):** the entire game runs on **Qwen2.5-1.5B** — ~1.6B total runtime | |
| params (LLM + Supertonic), far under the 32B cap. | |
| - **Field Notes:** [docs/FIELD_NOTES.md](docs/FIELD_NOTES.md) — what we built, what | |
| broke, and what a 1.5B model can really do. | |
| - **Sharing is Caring:** real agent traces (case generation + live interrogations) are | |
| published on the Hub — see the traces dataset linked below. | |
| See [COMPLIANCE.md](COMPLIANCE.md) for the full parameter budget and badge details. | |
| ## 📦 Submission links | |
| - **Demo video:** *coming before the deadline — link will land here.* | |
| - **Social post:** *coming before the deadline — link will land here.* | |
| - **Agent traces dataset:** [build-small-hackathon/case0-traces](https://huggingface.co/datasets/build-small-hackathon/case0-traces) | |
| — the exact prompts + completions of a full case generation, and a live 4-turn interrogation. | |
| ## ▶️ Run it locally | |
| ```bash | |
| # 1. backend deps + open weights | |
| python -m venv .venv && .venv/Scripts/pip install -r requirements.txt # (Windows) | |
| python scripts/fetch_models.py # one-time: fetch the open GGUF + Supertonic | |
| # 2. build the pixel-art frontend bundle (served by gradio.Server from web/dist) | |
| cd web && npm install && npm run build && cd .. | |
| # 3. run — open http://127.0.0.1:7860 | |
| python app.py | |
| ``` | |
| The game runs entirely on the CPU — laptop or Space, same code, no GPU required. | |
| (In the Docker/Space build both steps happen automatically: a Node stage builds the | |
| bundle and the Python stage compiles llama.cpp and bakes the weights.) | |
| ## 🙏 Credits | |
| - **LLM:** [Qwen2.5-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct) (Apache-2.0), via llama.cpp. | |
| - **Voices:** Supertonic on-device TTS. | |
| - **Music:** *"Backbay Lounge"* by Kevin MacLeod (incompetech.com), licensed under | |
| [Creative Commons Attribution 4.0](https://creativecommons.org/licenses/by/4.0/). | |
| - **Fonts:** Silkscreen & Pixelify Sans (SIL Open Font License), self-hosted. | |
| - Pixel art and UI sound effects: procedurally generated. |