arbios/mindlock-minicpmv46-departments-lora

mAIndlock — Department LoRA (MiniCPM-V 4.6)

A small LoRA adapter that distils the sensory-region behavior of the NPC brains in mAIndlock into MiniCPM-V 4.6.

In mAIndlock every character's decision is computed by six roles drawn from decision neuroscience — four of them sensory regions that each rate one dimension of the player's words: amygdala (threat), hippocampus (memory + trust/fear lean), striatum (habitual reward), ACC (effort / cost). A base small model tends to flatten these — it answers a sincere plea and a veiled threat almost the same way. This adapter is trained to pull them apart, so the regions react sharply and differently to cruelty vs. sincerity.

• Base: openbmb/MiniCPM-V-4.6
• Method: LoRA (rank 16, α 32), ms-swift, bf16, 3 epochs, A100 (Modal)
• Data: a few hundred role-specific examples — system prompt per region, one player line in, one structured rating out

from peft import PeftModel
from transformers import AutoModel
base = AutoModel.from_pretrained("openbmb/MiniCPM-V-4.6", trust_remote_code=True)
model = PeftModel.from_pretrained(base, "arbios/mindlock-minicpmv46-departments-lora")

The live Space runs the quantized MiniCPM-1B regions on llama.cpp; this adapter is the fine-tuned counterpart that demonstrates department differentiation (the project's "before/after").

• Project / code: https://github.com/Arbios/maindlock
• How the brain works: https://github.com/Arbios/maindlock/blob/main/docs/ARCHITECTURE.md
• Deliberation traces: https://huggingface.co/datasets/arbios/maindlock-brain-traces

License follows the base model, openbmb/MiniCPM-V-4.6. Built for the Hugging Face × Gradio "Build Small" hackathon.