Planning allowance: 3–10 working days after data preparation. Choose a base model that already handles the task reasonably and whose licence fits the intended use. Pin the checkpoint revision, tokenizer, template, package versions and random seeds.
Hugging Face’s SFT Trainer accepts supported dataset formats; PEFT’s quantisation guide explains adapting quantised models with trainable adapters. Use the current documentation for implementation, and keep a dependency lockfile. This plan does not claim that an arbitrary package version or GPU will run every model.
Start with a small dataset and limited configuration sweep. Log trained tokens, validation passes, wall time, failed jobs and human effort. Check that target text and loss masking match the task. Inspect truncated inputs and outputs; a low training loss can conceal unusable examples.
Compare at least the unchanged base model and the adapted model. If resources allow, test a second data slice or adaptation configuration to distinguish a method effect from a lucky run. The LoRA learning/forgetting study motivates checking retained capabilities as well as the target task. Do not keep retuning against the final holdout.
Stop condition: improvement is confined to seen templates, labels cannot be trusted, or the proposed deployment cannot meet the operating constraints.