Fine-Tuning & LLMOps

When to fine-tune vs. RAG vs. prompt engineering: a sovereign developer's decision framework. Covers task-fit analysis, dataset requirements, compute planning, and expected outcomes.

Fine-tune 8B+ models on consumer GPUs: QLoRA with Unsloth, dataset preparation, training configuration, memory optimisation, and exporting to GGUF for sovereign local inference.

Evaluate sovereign LLM systems: LLM-as-judge frameworks, RAGAS for RAG evaluation, task-specific metrics, perplexity baseline comparison, and human evaluation workflows.

Sovereign LLM output safety: input/output validation with Guardrails AI and NeMo Guardrails, hallucination detection, toxicity filtering, and responsible deployment patterns.

Self-host LLM inference servers: Ollama API, llama.cpp server, vLLM for throughput, OpenAI-compatible endpoints, and sovereign serving behind Nginx with authentication.