Give your adaptation the capacity it needs.

Define the gain that will justify your run.

Define an observable task: classifying a ticket, extracting fields, or following a response format. First work on a reference instruction and keep its outputs. If a prompt change is enough, it becomes a candidate for comparison. Without this baseline, a gain attributed to fine-tuning may simply come from a better formulation or a different handling of responses.

Separate format validity from content correctness. A JSON object that passes the parser can contain the wrong category. Add the behaviors to preserve, for example how a missing field or a rare class is handled. Set the acceptance rules and rejection reasons before choosing the rank, learning rate, or number of steps.

Choose data that lets you decide.

Keep the variants of the same conversation, document, or source in a single set when their proximity threatens the independence of the evaluation. Make the split before learned transformations or result-guided tuning. A random split by line can give the impression of progress on examples that have already been almost seen.

Pedagogical example: 1,200 tickets grouped by conversation are divided into 840 training examples, 180 validation examples, and 180 test examples after respecting the groups. These numbers describe an illustrative plan, not a provided corpus. The validation set is used to choose the variant; the final test set is kept aside. Inspect the categories and lengths in each set rather than assuming that the ratio alone guarantees their quality.

Pick the adaptation to compare against your baseline.

Full fine-tuning modifies the relevant weights of the base. LoRA adds low-rank matrices to the targeted modules and learns their contribution. The rank, targeted modules, scale, and selected options define the experiment. A higher rank increases the number of adapter parameters without demonstrating that it will improve your task.

For an illustrative 4,096 × 4,096 matrix, two rank-8 LoRA matrices contain 8 × (4,096 + 4,096) = 65,536 values, versus 16,777,216 in the dense matrix, excluding biases and other modules. This ratio of numbers describes neither the whole model nor the memory peak: the base, activations, and learned parameter states still need to be considered.

Technical sources: PEFT — configuration and LoRA matrices

Choose the precision for the whole workflow.

A base loaded in four bits does not mean that all operations, all tensors, and the cache use four bits. Quantization metadata and modules kept in another precision influence the allocation. In bitsandbytes, training models loaded in four or eight bits concerns additional parameters; it is not a general update of the quantized weights.

Record the base format, the compute type, the trainable modules, and the versions. Compare a real step, the evaluation, and the export operation you want to use. Merging, dequantizing, or producing another format may require a different amount of space than training. A method that holds up during adaptation is not yet proof that the whole cycle will fit on the same card.

Technical sources: Transformers 5.17 — bitsandbytes and additional parameters

Compare your variants with a defined trial budget.

On the illustrative set, keep a baseline without adaptation, a rank-8 LoRA, and a rank-16 variant, with the same modules, data, step budget, and evaluation rules. The rank is then the factor being studied. If you change both rank and learning rate at the same time, you create another question; the ablations guide explains how to make that interaction readable.

Before each trial, check a few tokenized examples and the portions that contribute to the loss. After the trial, compare the identified predictions with the same evaluator. Note missing outputs and regressions, not just the average. A rare class that becomes systematically incorrect can invalidate an overall gain depending on your quality contract.

Also document alpha, dropout, and initialization. With the classic LoRA setting, the scale is alpha/r; state whether you keep alpha or the alpha/r ratio when the rank changes. These protocols do not answer exactly the same question. Another scaling rule, such as rsLoRA, must be identified separately.

Technical sources: PEFT — rank, alpha and LoRA scaling rules

Decide what to fix before committing more compute.

A loss that keeps falling while validation quality stalls calls for analysis, not automatically more steps. Examine label errors, repeated phrasing, truncated inputs and divergences between the training format and the inference format. Keep a sample of errors per category to choose the fix.

If you fix a taxonomy or the parsing, version the evaluator and recompute the scores of the candidates that are needed. Do not directly compare two numbers obtained with different rules. If you inspect the final test to choose a fix, it now takes part in development: prepare a new independent evaluation for the next conclusion.

Verify the result you want to reuse.

The PEFT format keeps the adapter parameters and its configuration; it generally does not contain the base weights. So keep the exact identifier and revision of that base alongside the artifact. A model name alone that points to an evolving version is not enough to reconstruct the conditions.

Reload in a new process the result actually intended for use. Replay a few control inputs, then the evaluation protocol. Archive the tokenizer, template, generation settings, dependencies, metrics and data provenance. If the artifact is merged, note that transformation and check its outputs; it is a step to verify, not an invisible formality.

Technical sources: PEFT — checkpoint contents and base model

Choose the next step from the results.

Compare GPUs based on the peak of the full step and the evaluation, the maximum inputs required and the compatible software stack. In the plan, set aside time for the control, the priority variants, reading the errors and recovery. Choosing an environment at order time expresses an intended preparation, without automatically qualifying your adaptation.

The conclusion must name the candidate retained or the absence of sufficient gain, the groups covered, the known regressions and the reload conditions. Your log keeps the references and the reason for the choice. The memory measurement pack helps understand the instrumentation; it does not provide a LoRA training tested in your place.