--- title: Two Truths & a Lie — 20Q emoji: 🐨 colorFrom: yellow colorTo: blue sdk: gradio sdk_version: 6.19.0 python_version: '3.12' app_file: app.py pinned: false license: apache-2.0 short_description: Lie to a koala detective's face. See if you walk out. models: - prism-ml/Ternary-Bonsai-8B-gguf - prism-ml/Ternary-Bonsai-27B-gguf --- # 🐨 TWO TRUTHS & A LIE — An Interrogation in Twenty Questions **Lie to a detective's face. See if you walk out.** You write three short claims about yourself — two true, one a lie — and take a seat in Interview Room 3 of the Gumleaf Precinct. A koala detective asks you up to twenty short-answer questions, then accuses. If the accusation is wrong, you fooled the machine. Who's on duty tonight: - **☕ Det. Pip Barrow** — Ternary-Bonsai-**8B** (1.58-bit). Fast. Jumpy. Guesses. - **🌧️ Insp. Marlowe Grey** — Ternary-Bonsai-**27B** (1.58-bit). Slow to wake. Hard to fool. Both run on **ZeroGPU** through the PrismML llama.cpp fork (vendored in `./bin`), one per-call `llama-server` per turn. ## 🔴 The tape is rolling — data collection, plainly Every interrogation is **recorded and used as AI training/evaluation data**: your three claims, the hidden lie label, the full Q&A transcript, the detective's accusation, and the observation-room analysis. Records are stored as JSONL in a **private** dataset repo. No account and no name are attached — only what you type. A persistent 🔴 REC banner and a consent checkbox gate every game; if you don't consent, nothing runs and nothing is saved. One more honest detail, revealed in-game after each case: a **second, hidden model instance** ("Sgt. Wattle, observation room") reads your answers as you give them and slips behavioral notes to the detective between questions. You never see it during play — you do see its full notes in your case file afterward. ## Running locally ```bash pip install -r requirements.txt MOCK_LLM=1 python app.py # scripted brains, no GPU needed python tests/drive_mock.py # full game-loop test suite (mock mode) ``` Env knobs: `DATASET_REPO` (default `devmandan/2t1l-games`), `DISABLE_PUSH=1`, `SKIP_27B=1`, `MOCK_LLM=1`.