Fine-tuned Gemma 4 E4B drives 100 villagers offline in a rumor-spread simulation game
AlternativeEnough352 · reddit · 2026-11-23
A developer open-sourced Rumor Village: a 3D village where up to 100 neighbors are driven by a fine-tuned Gemma 4 E4B running locally via llama.cpp — fully offline, no API. Each simulated minute, every villager submits a 256-token card and outputs one GBNF-constrained JSON action; a C++ engine validates actions against memory ("the model proposes, the engine decides") with exact integer-math replay. Fine-tuning cut fabricated memories from 2% to 0%, reduced failures 5–16.5x, and halved generation time to 0.36s/min at 4.4–4.6GB VRAM on an RTX 5070 Ti (rejected actions rose to 1%). Trained with LoRA (r16) on 12,146 bilingual card→answer pairs generated by two frontier models and engine-validated. MIT code, Apache-2.0 model, plus a GPU-free web demo.
More from coding & agent
- Apple cuts per-dev daily token budget from $300 to $100; cheaper-model strategies follow — Zachly · 2026-11-23
- A lightweight script makes CLI agents debate each other to a consensus write-up — reddit_kwr · 2026-11-23
- Nous ships official Hermes Blender plugin: agents drive live 3D scenes via MCP — Teknium · 2026-11-23
- Dev builds Battlefield 1942-style browser FPS with GPT-6.2: 16v16 battles, vehicles, AI squads — VraserX · 2026-11-23
- Dev spins up hundreds of AI agents in a livestreamed attempt to solve P = NP — amantinband · 2026-11-23
- A 9-machine local AI homelab: 5 agents on one dual-3090 llama-server — StandardLovers · 2026-11-23