README.md

metadata

license: mit
language:
  - en
pipeline_tag: text-generation
library_name: transformers
tags:
  - Modular
  - Intelligence

Modular Intelligence

A Small but Complete Reasoning System Built on Top of a Language Model

System-Level View (60 Seconds)

This repository is not a standard GPT-2 model.
It is a compact reasoning system built on top of an LLM, using:

• Modules – structured reasoning lenses
• Checkers – strict second-pass reviewers
• Contracts – fixed output sections for every module
• Optional Router – automatic module selection from free-form tasks

Pipeline:

Task → Lens (Module) → Structured Output → Checker (Optional)

The LLM engine is replaceable; the architecture is the system.

---

Architecture

+—————————————————————+ | MODULAR INTELLIGENCE | +—————————————————————+ | User Task | | | | | v | | [Auto-router or manual module selection] | | | | | v | | +———————–+ +————————+ | | | GENERATOR MODULE | ––> | CHECKER MODULE | | | | (structured output) | | (optional verification)| | | +———————–+ +————————+ | | | ^ | | v | | | Structured Output Feedback / Fix Ideas | +—————————————————————+

---

Repository Contents

• modules.json — all modules, inputs, sections, checkers
• app.py — Gradio UI with optional auto-routing
• Model card
• GPT-2 base model (fully swappable)

---

Problem This Solves

Standard LLMs: unstructured text, drift, no audit trail.
Modular Intelligence: task decomposition, structured reasoning, verification, consistency, interpretability.

---

Modules (Lenses)

Module	Use
Analysis Note	Explain or break down text/situations
Document Explainer	Summaries of contracts/policies/articles
Strategy Memo	Options → Recommendation → Risks → Next Steps
Message/Post Reply	Structured replies
Profile/Application	Bios, cover letters, statements
System Blueprint	Design or improve systems and workflows
Modular Brainstorm	Decompose problems into modules/checkers

---

Checker Layer

Checkers produce:

• VERDICT
• ALIGNMENT
• STRUCTURE
• GAPS
• FIXES

Implements: Reason → Critique → Refine.

---

End-to-End Example

Task: “Should we expand into a new city in 6 months?”

1. Select strategy_memo_v1
2. Fill: context, objective, constraints
3. Run → structured memo with required sections
4. Run checker → verdict + issues + fixes

---

Auto-Routing (Optional)

The provided app.py includes an Auto-Route tab:

1. Paste free-form task
2. Zero-shot classifier ranks modules
3. Jump directly to best lens

---

Code Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

tok = AutoTokenizer.from_pretrained("botbottingbot/Modular_Intelligence")
model = AutoModelForCausalLM.from_pretrained("botbottingbot/Modular_Intelligence")

prompt = """
MODULE: Strategy Memo
INPUTS:
CONTEXT: We are expanding to City X.
OBJECTIVE: Decide whether to enter within 6 months.
CONSTRAINTS: Budget limits; regulatory uncertainty.

OUTPUT SECTIONS:
- CONTEXT
- OBJECTIVE
- CONSTRAINTS
- OPTIONS
- RECOMMENDATION
- RISKS
- NEXT_ACTIONS
"""

inp = tok(prompt, return_tensors="pt")
out = model.generate(**inp, max_new_tokens=250)
print(tok.decode(out[0], skip_special_tokens=