Buzzwords & Misdemeanors: a field log of teaching one 1B model to run a courtroom
BastienHot
โข โข 1A single multitask LoRA on MiniCPM5-1B that plays the Game Master: writes the hidden Case File, directs the beat loop (GBNF-constrained JSON), and scores the player guess. Apply on the MiniCPM5-1B GGUF via llama.cpp lora_path. Distilled from Gemma teacher traces โ see github.com/BastienHot/BuildSmallHackathon (training/director_*.py).
We're not able to determine the quantization variants.
Base model
openbmb/MiniCPM5-1B