Fine-Tuning jobs

Fine-tuning roles live in the gap between prompting and pretraining: adapting open-weight models — Llama, Qwen, Mistral — to a domain, a format, or a latency budget the API models can't hit. Dataset construction comes first, training second, and third comes proving the tuned model beats a well-prompted frontier model, because that comparison decides whether the project survives.

LoRA and QLoRA via Axolotl, Unsloth, or raw Hugging Face are the standard moves, with DPO and its cousins for preference tuning and quantization for deployment. Knowing when not to fine-tune is a screening question. Roles concentrate where proprietary data or unit economics justify the complexity — the only two reasons that hold up.