What problem does it solve? Teams reflexively jump to fine-tuning (or avoid it entirely) without evidence, wasting GPU budget or shipping inconsistent model behavior. This Skill provides a measured decision framework across the four adaptation levers and a reproducible PEFT training workflow that keeps proprietary data on self-hosted GPUs. ## Core Features & Use Cases - Adaptation decision framework: Compares prompt engineering, RAG, PEFT/LoRA, and full fine-tuning across data volume, knowledge drift, latency economics, IP constraints, and ops maturity, with a break-even token-cost analysis. - PEFT workflow design: Covers LoRA/AdaLoRA/QLoRA configuration, response-masked supervised datasets, 4-bit quantized training with Hugging Face Trainer, and trainable-parameter sanity checks. - Kubernetes training jobs: Specifies a K8s Job with pinned CUDA images, secrets management, and checkpointing to S3-compatible object storage so preemptible GPU nodes can resume training. - Dual evaluation protocol: Combines ROUGE, embedding semantic similarity, consistency probes, and blind domain-expert review against few-shot baselines before any ship decision. - Use Case: A research team wants a domain assistant with consistent house voice. The Skill baselines few-shot prompting, decides PEFT is the right lever, builds a JSONL dataset, runs a QLoRA Job on a single GPU with checkpoint resume, and evaluates the adapter against the baseline before MLflow registration. ## Quick Start Use the finetuning-strategy-advisor skill to decide whether I should fine-tune or use RAG for my domain Q&A assistant and design the training job.
