neuron-one ML Model Training
Pipeline
- Define task and success metrics
- Collect/prepare training data
- Choose base model (by size and task fit)
- Select GPU and estimate cost
- Configure training (SFT, DPO, or GRPO)
- Monitor training run
- Evaluate on test set
- Deploy model
Model Selection
- Small tasks (classification): Qwen 0.6B-3B
- Medium tasks (generation): Mistral 7B, Llama 8B
- Complex tasks (reasoning): Qwen 27B, Llama 70B
Training Methods
- SFT: supervised fine-tuning on examples
- DPO: preference optimization (good vs bad)
- GRPO: group relative policy optimization
Deployment
- Ollama for local inference
- LitServe for MCP server
- llama.cpp for production