LARS - Local AI Runtime System
Base Model Options
- Qwen 2.5 Abliterated - Current LARS base (7B-14B)
- Qwen 3 Coder - For coding-focused deployments
- Custom bases - Client-specific requirements
Training Stack
- Hardware: RTX 3090s / RTX 4090s / A100s
- Framework: Transformers + PEFT (LoRA)
- Location: Local AI server (no cloud training)
Training Types
Identity Training
- Who is LARS?
- Relationship to Nexus
- Client-specific customization
- Personality and tone
Knowledge Training
- Client documents and data
- Domain expertise
- Procedures and workflows
- Historical context
Reasoning Training (3D Dataset)
- Thinking modes
- Multi-step problem solving
- Context evaluation
- Decision frameworks
Verified Training Loops
Not one-shot fine-tuning. Continuous loop: 1. Train on dataset 2. Test with evaluation suite 3. Score with AI judge (Claude) 4. Generate corrections for failures 5. Retrain with corrections 6. Loop until 98%+ accuracy
Model Storage
- LoRA adapters for efficient switching
- Multiple versions (identity, knowledge, reasoning)
- Merge capabilities for final deployment
- GGUF conversion for Ollama deployment