Open AI Box – Universal LLM introspection: injection points & dimension roles in any model

Hey everyone

I’m releasing Open AI Box — a Python package that opens the black box of any LLM.

What it does

By tracing a live inference pass with PyTorch forward hooks, it automatically discovers:

• Injection points — where data enters, where decisions are made, where memory lives

• Dimension roles — which dimensions carry causality, emotion, certainty, temporal reasoning

• Everything exported to a single graph.json

Live demo

Open AI Box - a Hugging Face Space by anzizdaouda0

Paste any model ID → get a full architecture analysis with interactive Plotly graphs.

Install

pip install openaibox

from openaibox import OpenAIBox

oaib = OpenAIBox(“HuggingFaceTB/SmolLM-360M”)

oaib.discover().map_dimensions().export(“graph.json”)

oaib.print_summary()

Key finding: the model.norm → lm_head control point is universal across all tested architectures (LlamaForCausalLM, Qwen2ForCausalLM, GPT2…).

PyPI: openaibox · PyPI

GitHub: GitHub - Tryboy869/openaibox: Universal LLM introspection. Open AI Box — understand any model. · GitHub

Happy to hear feedback!