📺 Watch the full tutorial on YouTube

🧠 TwIL-LM3 Chain-of-Thought Visualizer

A specialized 3B parameter reasoning engine from webAI Intelligence Lab that outreasons OpenAI's 120B gpt-oss model on 4 of 5 formal reasoning benchmarks while running locally on edge consumer hardware (from an iPhone to a Raspberry Pi).

⚡ 3-Step Workflow

Step 1: Input 📥	Step 2: AI Action ⚙️	Step 3: Result 📊
Supply formal logic rules, policy text, or premise statements.	TwIL engine generates <think> reasoning traces and validates logic steps.	Produces side-by-side comparison report with step metrics and answer accuracy.

---

🌟 Primary Model Powers & Highlights

• 40x Parameter Efficiency: Outperforms OpenAI's gpt-oss-120b (120B params) on 4/5 formal reasoning benchmarks at just 3B parameters.
• 2.6x Faster Inference: Delivers 2.6x faster completed answer throughput compared to 120B frontier models.
• Proprietary Verified Datasets: Trained on webAI-owned, verified logic corpora rather than raw scraped web text.
• Edge & Mobile Native: Designed to execute fully offline on consumer hardware, from iPhones and Raspberry Pis to local workstations.
• Specialized Deductive Reasoning: Built specifically for formal logic, rule induction, contract auditing, tool calling, and AI agent verification.

---

🛠️ Technology Stack

• Core Engine: Python 3.10+
• Inference Runtime: Local Ollama API (twil-lm3:3b) / Hugging Face transformers & llama.cpp
• Target Model: webAI-Official/TwIL-LM3 (SmolLM3-3B Reasoning Derivative)
• Output Report: Structured Markdown Benchmarks (outputs.md)

---

📂 Project Structure

WebAI Twil/
├── README.md                          # Comprehensive user documentation & setup guide
├── main.py                            # Primary execution pipeline & logic benchmark engine
├── download_and_register_twil.py      # Automated script to download weights & register model in Ollama
└── README (94).md                     # Official Hugging Face model specification sheet

---

📥 How to Download & Register TwIL-LM3 Model in Ollama

You can easily register twil-lm3:3b in your local Ollama runtime whenever you are ready:

Option A: Automated Python Registration Script

Run the included download script to fetch TwIL-LM3-Q4_K_M.gguf (1.91 GB) from Hugging Face and register it automatically:

python download_and_register_twil.py

Option B: Manual Command-Line Registration

1. Download TwIL-LM3-Q4_K_M.gguf directly from webAI-Official/TwIL-LM3 on Hugging Face.
2. Save the GGUF file in your project folder.
3. Create a Modelfile with the following contents:

   FROM ./TwIL-LM3-Q4_K_M.gguf
   PARAMETER stop "<|im_end|>"
   PARAMETER stop "<|endoftext|>"

4. Register the model in Ollama:

   ollama create twil-lm3:3b -f Modelfile

---

📱 Edge Device Deployment (iPhone & Raspberry Pi)

1. iPhone (iOS Setup)

TwIL-LM3's compact GGUF quantization (1.78 GB) runs locally on iPhones without cloud connections:

1. Install App: Download PocketPal AI or MLC LLM from the iOS App Store.
2. Download Weights: Fetch TwIL-LM3-Q4_K_M.gguf via Safari or transfer via AirDrop.
3. Import Model: Open PocketPal AI -> Models -> Import Local Model and select TwIL-LM3-Q4_K_M.gguf.
4. Set Generation Budget: Set max tokens to 2048 and temperature to 0 (greedy decoding) so <think> reasoning blocks do not truncate.

2. Raspberry Pi (SBC Setup)

Run TwIL-LM3 on a 4GB+ Raspberry Pi 4 or 5:

1. Install Build Tools:

   sudo apt update && sudo apt install -y git build-essential cmake

2. Compile llama.cpp:

   git clone https://github.com/ggerganov/llama.cpp
   cd llama.cpp && cmake -B build && cmake --build build --config Release

3. Download Model File:

   wget https://huggingface.co/webAI-Official/TwIL-LM3/resolve/main/TwIL-LM3-Q4_K_M.gguf

4. Run Local Inference:

   ./build/bin/llama-cli -m TwIL-LM3-Q4_K_M.gguf -cnv --temp 0 -n 2048

---

🚀 Running the Reasoning Visualizer

Once twil-lm3:3b is registered, execute main.py to process the formal logic benchmark suite:

python main.py

After execution, open outputs.md to inspect side-by-side deductive reasoning steps (<think> traces).

---

💡 Real-World Use Cases

1. Enterprise Policy Compliance: Audits business guidelines and contract rules for logical contradictions.
2. First-Order Logic Parsing: Converts natural language premises into verifiable symbolic predicate expressions.
3. Rule Induction & Derivation: Deduces underlying constraint rules from state transition pairs.
4. Lean Math Formalization: Verifies mathematical proofs step-by-step prior to theorem prover input.
5. High-Speed Agent Verification: Validates tool-calling decisions and agent safety rules with zero cloud dependency.

---

🔮 Future Features

• Lean Proof Auto-Corrector: Fixes syntax flaws detected in reasoning traces automatically.
• CLI Streaming Dashboard: Live terminal stream of step-by-step <think> block deductions.
• Multi-Contract Loophole Auditor: Scans multi-page legal documents for conflicting clauses.
• Mobile GGUF Profiler: Measures latency and VRAM footprint on mobile edge hardware.
• Symbolic Graph Exporter: Converts textual deductive proofs into interactive visual flowcharts.

---

🏷️ Keywords & Search Tags

webAI TwIL-LM3 Formal Logic AI Edge AI iPhone Local LLM Raspberry Pi AI Reasoning Model Chain of Thought GGUF Quantization SmolLM3 Local LLM Deductive Verification