Ulshe EasyBake AI Forge v1.0
ulshe_ai
HCTS Easy Bake AI: Mobile Forge & Dojo
Welcome to the HCTS Easy Bake AI mobile application. This software represents a breakthrough in decentralized artificial intelligence, allowing users to train, refine, and deploy high-fidelity Transformer models directly on mobile hardware.
The app utilizes the Forge Engine—a custom training pipeline designed to bypass the traditional need for massive cloud compute by leveraging a staged curriculum and local optimization.
🚀 The Forge: Phased Curriculum Training
Unlike standard AI training that attempts to learn everything at once, the Easy Bake Forge uses a 9-Phase Genesis Curriculum. This approach mirrors human cognitive development, moving from simple primitives to final synthesis.
The 9 Phases:
Phase 0-1: Primitives & Lexicon: The AI learns basic character relationships and builds its foundational vocabulary.
Phase 2-3: Linguistics & Logic: Focuses on syntax, sentence structure, and basic reasoning patterns.
Phase 4-5: Math & Ethics: Introduces numerical systems and the Core Ontology/Ethical alignment.
Phase 6-7: Physics & Social: Implements understanding of physical constants and interactive social dialogue.
Phase 8: Final Synthesis: Aggregates all previous data for a unified, high-reasoning intelligence core.
🥋 The Reinforcement Dojo
The Dojo allows for Direct Logic Intervention. If the model responds incorrectly to a specific prompt, you can administer a Logic Jolt.
Manual Jolt: Train the model on a single Question/Answer pair to correct a specific behavior.
Automated Mastery: Load a .jsonl dataset, and the Dojo will iteratively test and re-train the model until every card in the set is mastered (100% accuracy).
Telemetry: Watch the Shift Delta to see the precise numerical change in the model's internal weights.
⚙️ Understanding the Specs (Advanced Settings)
d_model (The Width)
Think of d_model as the Resolution of the AI's internal thoughts.
512+: High definition. Capable of complex nuance but requires significant RAM.
256: Standard mobile performance. Balanced and stable.
128: Low-res logic. Extremely fast, but the AI may struggle with complex linguistics.
Layers (The Depth)
The number of Encoder and Decoder layers defines the AI's Chain of Thought. More layers allow the model to process information through more steps of reasoning.
Attention Heads
These are the parallel "eyes" of the model. They allow the AI to look at multiple parts of a sentence simultaneously to understand context.
🛠️ Performance & RAM Management
Mobile devices have limited memory compared to servers. If the app crashes during the "Forge" process:
Reduce Batch Size: Drop from 32 to 8 or 4. This is the most effective way to save RAM.
Lower d_model: If your device has < 6GB of RAM, use 256 or lower.
Shorten Max Seq Length: Reducing this from 256 to 128 significantly lightens the computational load.
Battery Optimization: Set the app to "Unrestricted" in Android Settings to prevent the OS from killing the Forge engine while it's running in the background.
📂 Model Storage
All trained cores (.pth files) and datasets (.jsonl) are saved to your public device storage for easy extraction: Internal Storage > Download > UlsheForgeModels
© 2026 Ulshe AI - Intelligence Forged Locally.
Welcome to the HCTS Easy Bake AI mobile application. This software represents a breakthrough in decentralized artificial intelligence, allowing users to train, refine, and deploy high-fidelity Transformer models directly on mobile hardware.
The app utilizes the Forge Engine—a custom training pipeline designed to bypass the traditional need for massive cloud compute by leveraging a staged curriculum and local optimization.
🚀 The Forge: Phased Curriculum Training
Unlike standard AI training that attempts to learn everything at once, the Easy Bake Forge uses a 9-Phase Genesis Curriculum. This approach mirrors human cognitive development, moving from simple primitives to final synthesis.
The 9 Phases:
Phase 0-1: Primitives & Lexicon: The AI learns basic character relationships and builds its foundational vocabulary.
Phase 2-3: Linguistics & Logic: Focuses on syntax, sentence structure, and basic reasoning patterns.
Phase 4-5: Math & Ethics: Introduces numerical systems and the Core Ontology/Ethical alignment.
Phase 6-7: Physics & Social: Implements understanding of physical constants and interactive social dialogue.
Phase 8: Final Synthesis: Aggregates all previous data for a unified, high-reasoning intelligence core.
🥋 The Reinforcement Dojo
The Dojo allows for Direct Logic Intervention. If the model responds incorrectly to a specific prompt, you can administer a Logic Jolt.
Manual Jolt: Train the model on a single Question/Answer pair to correct a specific behavior.
Automated Mastery: Load a .jsonl dataset, and the Dojo will iteratively test and re-train the model until every card in the set is mastered (100% accuracy).
Telemetry: Watch the Shift Delta to see the precise numerical change in the model's internal weights.
⚙️ Understanding the Specs (Advanced Settings)
d_model (The Width)
Think of d_model as the Resolution of the AI's internal thoughts.
512+: High definition. Capable of complex nuance but requires significant RAM.
256: Standard mobile performance. Balanced and stable.
128: Low-res logic. Extremely fast, but the AI may struggle with complex linguistics.
Layers (The Depth)
The number of Encoder and Decoder layers defines the AI's Chain of Thought. More layers allow the model to process information through more steps of reasoning.
Attention Heads
These are the parallel "eyes" of the model. They allow the AI to look at multiple parts of a sentence simultaneously to understand context.
🛠️ Performance & RAM Management
Mobile devices have limited memory compared to servers. If the app crashes during the "Forge" process:
Reduce Batch Size: Drop from 32 to 8 or 4. This is the most effective way to save RAM.
Lower d_model: If your device has < 6GB of RAM, use 256 or lower.
Shorten Max Seq Length: Reducing this from 256 to 128 significantly lightens the computational load.
Battery Optimization: Set the app to "Unrestricted" in Android Settings to prevent the OS from killing the Forge engine while it's running in the background.
📂 Model Storage
All trained cores (.pth files) and datasets (.jsonl) are saved to your public device storage for easy extraction: Internal Storage > Download > UlsheForgeModels
© 2026 Ulshe AI - Intelligence Forged Locally.
Developer email
joreag@ulshe.ai
Developer website
https://www.ulshe.aiReleased
Aug 11, 2026
Current version
1.1
Android version
8.0
Screenshots
