Spaces:
Sleeping
Sleeping
| 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 | |
| 🙌 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. | |