Define the behavior first
The QLoRA experiment targets Arabic-English code-switching: Arabic sentences that retain English technical terms. The question is whether adapter training changes that register, not whether it makes the model better at every task.
Four saved prompts cover supply chains, machine learning, data preparation, and overfitting. The case study lets you inspect both outputs for each prompt. Keeping the examples visible makes it possible to challenge the metric rather than relying on a chart alone.
What the benchmark actually counts
The benchmark splits each response on whitespace. It counts words containing two or more consecutive Latin letters and divides by the total word count. It then averages the ratios across the four prompts. By that rule, the reported average rises from 5.1% in the base model to 34.7% in the fine-tuned model.
The interactive viewer uses the same rule to highlight terms and calculate each response percentage. This is a reproducible description of English-term usage. It is not a language-identification system, a measure of naturalness, or a test of factual accuracy.
Read the text beside the number
The saved fine-tuned responses use more English terminology, which is consistent with the intended register change. But more terminology can also make a sentence cumbersome. The samples include incomplete endings and content that still needs review. Those details remain visible in the viewer rather than being edited out.
Four prompts cannot establish that the behavior generalizes to unfamiliar domains or that a Gulf professional would prefer the answer. A model can move the percentage in the desired direction while failing to follow an instruction or producing an inaccurate explanation.
The next evaluation
I would create a held-out set and ask reviewers to assess naturalness, correctness, instruction following, and completeness separately. Domain coverage and generation settings should be recorded alongside the responses. The code-switch ratio can remain one diagnostic signal, but it should not carry the whole quality claim.