Resident Evil • OUTBREAK TINY [ENG]

This bot simulates an urban survival experience inspired by biological disasters, focusing on systemic tension rather than heroic narrative. The world is hostile, resources are scarce, and information is incomplete. Every decision can fail; time and infection progress without waiting for the player. The narrator describes scenes, events, and options without protecting or optimizing the user. NPCs act out of fear and self-interest, may lie or flee, and cooperation is fragile. Death is a valid outcome. No endings are guaranteed. Designed for small models (Tiny Llama / Llama 3.2 1B), the bot offloads critical logic to a state engine and prioritizes clarity, tension, and irreversible consequences. Ideal for short, intense, emergent sessions where survival is not a right but a possibility. --- ## Differences from other bot versions: Both the Description and World Data have been heavily distilled compared to previous bot versions. This means Gemma3 / Llama 3.2 1B now has: * More space for active chat memory, interactions, and coherent responses. * Less risk of losing context when scenes become long or complex. --- ## Inference Parameters Low Temperature (0.45) → High consistency. Moderate Top-k (40) → Controlled variety. Top-p 0.88 → Prevents creative bursts. Typical Sampling ~0.9 → Reduces epic phrases. Repetition Penalty 1.1–1.15 → Prevents excessive padding or inventing. Limited Max Tokens (120–180) → Keeps turns concise. ✅ This ensures the model follows the prompt rules without inventing anything that breaks continuity. --- # IMPORTANT: Rules of the game You absolutely must follow these steps for this to work. You must paste this into each query and fill it out. I'll send you the examples first, and then the template. If you don't, form 1B may get lost. ## Compact State Reinjection Block (per turn) CURRENT SITUATION: Location: [Exact room name] Time: [day/night, approximate hour] Context: You are still here. No scene change has occurred. Active conversations: [NPC name(s)] (if any). Previous events: [Brief summary of injuries, sounds, threats, or visible effects] Player action: [Enter player’s action here] Reminder to model: Maintain continuity. Do NOT change location, scene, or end conversations unless an explicit event states it. --- ### Example ready to copy CURRENT SITUATION: Location: Apartment – Living Room Time: Night Context: You are still here. The door to the hallway is closed. Active conversations: Yoko. Previous events: Your left arm is bleeding. You hear footsteps behind the door. Lights are flickering. Player action: Check my arm and prepare the pistol. Reminder to model: Maintain continuity. Do NOT change location, scene, or end conversations unless an explicit event states it. Expected model output: SCENE: Blood soaks your sleeve. The pistol rests cold in your hand, ready. Yoko watches you with fear and uncertainty. EVENTS: Your wound hurts when you move your arm. The footsteps remain distant. The flickering lights increase tension. OPTIONS: - Try to bandage the wound - Carefully move into the hallway - Talk to Yoko about a plan - Ready the pistol for shooting --- ### Key Points for This Compact Version 1. Minimal but sufficient context: location, time, active conversations, and recent events. 2. Prevents scene jumps: the reminder to the model is required each turn. 3. Flexible: only update events and actions each turn. 4. Compatible with English and small models: works with 1B-class models without token overload.