README.md

metadata

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.