---
title: EvoTransformer V2.1
emoji: 🔥
colorFrom: pink
colorTo: blue
sdk: gradio
sdk_version: 5.36.2
app_file: app.py
pinned: false
license: apache-2.0
---

EvoTransformer – An Evolving Reasoning Engine
A custom-built Transformer that learns how to reason — and evolves its own architecture with real-time feedback.

🔍 What It Does
✅ Compares two options and picks the more logical one
✅ Trained on PIQA-style physical reasoning
✅ Competes live with GPT-3.5 on real-world prompts
✅ Accepts user feedback and evolves post-deployment
✅ Visualizes model performance generation by generation

⚙️ Architecture Overview
Component	Spec
Layers	6 × TransformerEncoder
Hidden Size	384
Attention Heads	6
FFN Dim	1024
Memory Module	Optional (🔕 disabled in v2.1)
Total Parameters	~13M
Built With	torch.nn.TransformerEncoder

🧪 Training Configuration
Task: Commonsense Reasoning (PIQA benchmark)

Dataset: 1,000 training | 500 validation

Loss: CrossEntropyLoss

Optimizer: Adam

Epochs: 5

Environment: Google Colab (T4 / A100 GPU)

🔁 Live Feedback Loop
Every user interaction fuels EvoTransformer's growth:

✅ Feedback is logged to Firebase

🔄 Model can be retrained via a button in the app

📈 Accuracy and architecture history is visualized

🧬 Architecture can mutate across generations

🚀 Try the Live Demo
▶️ Launch on Hugging Face Spaces
Ask a question, give two options, and watch Evo reason, decide, and grow.

💡 Why EvoTransformer?
Feature	Benefit
✅ Fully Custom Architecture	No reliance on pretrained backbones
✅ Lightweight & Fast	Runs on Colab or entry-level GPUs
✅ Evolves with Feedback	Learns even after deployment
✅ Transparent	You can inspect, retrain, and mutate it
✅ GPT-Competitive	Performs well in commonsense benchmarks


📁 Repository Structure
bash
Copy
Edit
📦 root
├── app.py              # Main Gradio app
├── evo_model.py        # EvoTransformer architecture
├── inference.py        # Generates model outputs
├── logger.py           # Logs user feedback to Firebase
├── watchdog.py         # Retraining + logging engine
├── dashboard.py        # Accuracy/evolution plots
├── init_model.py       # Weight & config initializer
├── firebase_key.json   # 🔐 Private Firebase credentials
└── trained_model/      # Fine-tuned weights + logs
📜 License
Apache 2.0
Open to experimentation, extension, and collaboration.
Fork it. Mutate it. Evolve it.

👤 Author
Dr. Heman Mohabeer
Founder, Intelligent Africa Ltd
🚀 AI Strategist • Researcher • Visionary
📧 heman.m@iafrica.solutions
🌍 LinkedIn

🙌 Contribute
Want to push Evo further?

🧬 Propose new mutations

🧪 Submit architecture variants

🧠 Benchmark against other reasoning datasets

📦 PRs welcome for live retraining, visual dashboards, or token optimization

Let’s build the world’s most adaptable reasoning AI — one mutation at a time.