Universal LLM Introspection. Any model. Any architecture. Any provider.

Understand what happens inside any language model — without modifying it.

---

What is Open AI Box?

Open AI Box is a Python package that opens the black box of any LLM.

By tracing a live inference pass with PyTorch forward hooks, Open AI Box automatically discovers:

• The full architecture graph of the model
• The injection points — where data enters, where decisions are made, where memory lives
• The dimension roles — which of the hidden vector's dimensions carry causality, emotion, certainty, or temporal reasoning

Everything is exported to a single graph.json file.

---

Installation

pip install openaibox

Requirements: torch, transformers

---

Quickstart

from openaibox import Open AI Box

# Load and analyze any HuggingFace model
gr = Open AI Box("HuggingFaceTB/SmolLM-360M")

# Step 1 — Discover the architecture
gr.discover()

# Step 2 — Map dimension roles (optional, deeper analysis)
gr.map_dimensions()

# Step 3 — Export to graph.json
gr.export("graph.json")

# Step 4 — Print summary
gr.print_summary()

Output:

============================================================
  GRAPHRUNTIME — Analysis Report
  Model : HuggingFaceTB/SmolLM-360M
============================================================

  Architecture : LlamaForCausalLM
  Parameters   : 361,821,120
  Layers       : 30
  Hidden dim   : 960
  Vocabulary   : 49,152

  ──────────────────────────────────────────────────────
  INJECTION POINTS
  ──────────────────────────────────────────────────────

  [INPUT]     model.embed_tokens
  Shape  : [[1, 5]] → [[1, 5, 960]]
  Detail : Token embeddings — where the prompt enters the model.

  [DECISION]  model.norm
  Shape  : [[1, 1, 960]] → [[1, 1, 960]]
  Detail : Final normalization — the model's full understanding
           before deciding which token to output.

  [OUTPUT]    lm_head
  Shape  : [[1, 1, 960]] → [[1, 1, 49152]]
  Detail : Language model head — projects to vocabulary logits.

---

Core Concepts

Injection Points

Open AI Box identifies 4 types of injection points in any model:

Role	Description
input_point	Where the prompt enters (token embeddings)
decision_point	The richest reasoning state, just before token selection
memory_point	Where K/V projections encode context
output_point	Where logits over the vocabulary are computed

The decision_point is the most significant: it holds the model's complete understanding of the context before making a decision. Its shape is always [batch, seq, hidden_dim].

Dimension Map

Each dimension of the hidden vector carries specific information. Open AI Box identifies:

• Multi-role dimensions — active in many categories (global regulators)
• Specialist dimensions — active in only one category (precise carriers)

Example output for SmolLM-360M:

dim_696  →  6 roles  :  syntax, causality, certainty, abstraction, time, emotion
dim_544  →  6 roles  :  syntax, causality, certainty, abstraction, time, emotion
dim_792  →  5 roles  :  syntax, causality, certainty, time, emotion

Specialist dimensions:
  causality   : [295, 157]
  certainty   : [32, 545, 702, 683]
  temporality : [164, 93, 395]
  emotion     : [691, 950, 594, 644]

graph.json

The exported file contains the full analysis:

{
  "openaibox": {
    "version": "1.0.0b1",
    "model": "HuggingFaceTB/SmolLM-360M"
  },
  "architecture": {
    "class": "LlamaForCausalLM",
    "total_params": 361821120,
    "hidden_dim": 960,
    "vocab_size": 49152
  },
  "injection_points": [...],
  "layer_flow": [...],
  "dimension_map": {...}
}

See Understanding Results for a complete explanation of every field.

---

Supported Architectures

Open AI Box auto-detects the injection points for:

Family	Models
Llama	SmolLM, Llama-2/3, Mistral, Zephyr
Qwen	Qwen2, Qwen2.5, Qwen2.5-Math
GPT	GPT-2, CodeGen, StarCoder
BLOOM	BLOOM, BLOOMZ
Falcon	Falcon-7B/40B
OPT	OPT-125M → 66B
MPT	MPT-7B/30B
GPT-NeoX	Pythia, GPT-NeoX

Any model loaded from HuggingFace AutoModelForCausalLM is supported. If auto-detection fails, Open AI Box returns the full layer trace so you can identify the point manually.

---

Advanced Usage

Pre-loaded model

from transformers import AutoTokenizer, AutoModelForCausalLM
from openaibox import Open AI Box

model     = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1")
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")

gr = Open AI Box(model=model, tokenizer=tokenizer, model_name="Mistral-7B")
gr.discover().export("mistral_graph.json")

Custom contrast groups for dimension mapping

from openaibox.mapping import ContrastGroup
from openaibox import Open AI Box

my_groups = [
    ContrastGroup(
        name    = "code_vs_prose",
        label_a = "code",
        label_b = "prose",
        question = "Which dimensions carry code structure?",
        prompts_a = ["def add(a, b): return a + b", "for i in range(10): print(i)"],
        prompts_b = ["She walked slowly down the street.", "The sun set over the mountains."],
    )
]

gr = Open AI Box("microsoft/phi-2")
gr.discover()
gr.map_dimensions(groups=my_groups)
gr.export("phi2_code_analysis.json")

Explain a specific dimension

gr.discover().map_dimensions()
print(gr.explain_dimension(696))
# {
#   "dimension": 696,
#   "roles": ["syntax_semantics", "causality", "certainty", ...],
#   "global_score": 0.91,
#   "description": "Multi-role regulator (6 categories)"
# }

---

Creator

---

Documentation

• Understanding Results (EN)
• Comprendre les résultats (FR)
• README Français
• Changelog
• Code of Conduct

---

License

MIT — see LICENSE

---