mb8 0.1.0a0

Mathematical Brain 8-Level architecture — full training library

MB-8: Mathematical Brain 8-Level Architecture

MB-8 — full-training library for a novel 8-level neural architecture combining SSM, Hopfield networks, working memory, predictive coding, and plasticity mechanisms.

Features

• 8 architectural levels: Membrane → Synapse → Dendrites → SSM → Hopfield/Phase → Working Memory → Error → Plasticity/Consolidation
• Multiple model variants: SingleStream, DualStream, MB8ForCausalLM (full 8-level)
• Training modes: Pretrain (CLM), SFT (chat-style), GRPO (RL)
• Notebook API: Transformers-style Trainer + TrainingArguments for Jupyter/Colab
• CLI tools: mb8-train, mb8-sft, mb8-grpo, mb8-eval, mb8-inspect
• Epistemic framework: Hypothesis-driven development with go/no-go criteria

Quickstart

from mb8 import MB8Config, MB8ForCausalLM, TrainingArguments, Trainer

model = MB8ForCausalLM(MB8Config.from_preset("tiny_10m"))
trainer = Trainer(
    model=model,
    args=TrainingArguments(output_dir="outputs", max_steps=100),
    train_dataset=...,
)
trainer.train()

Installation

pip install mb8

Documentation

See api-docs/ for full reference and quickstart guide.

License

MIT