A legal document assistant built on a Foundry-hosted model needs to consistently output clauses using a strict numbering and formatting convention. The team has spent weeks refining the system message and adding detailed few-shot examples, but outputs still drift from the required format often enough to require manual correction. What should they try next?
Select an answer to reveal the explanation.
Short Explanation
Picture trying to teach someone a very specific paperwork format purely by handing them example after example every time they sit down to write, versus actually training them until the format becomes second nature. That is the difference between stacking more few-shot examples into an already long prompt and fine-tuning. The team here has already tried the first approach thoroughly, refining instructions and examples for weeks, and it still is not sticking reliably. That is the signal to move past showing examples and toward actually teaching the model the pattern directly, using a curated set of correctly formatted clauses so the behavior becomes built in rather than something it has to be reminded of on every call. Making responses longer does not fix a formatting problem, and dialing back safety filtering has nothing to do with clause numbering at all. When prompting alone has been pushed hard and still is not consistent enough, that is exactly the situation fine-tuning exists for.
Full Explanation
The correct answer is B. When extensive prompt engineering, including detailed few-shot examples, has already been tried and the model still cannot reliably reproduce a strict, specific output format, fine-tuning on a curated set of correctly formatted examples directly teaches the model that pattern, making the behavior far more consistent than instructions alone can achieve. Option A is incorrect because the problem described is formatting drift, not responses being truncated, so raising the output token limit does not address the actual failure mode. Option C is incorrect because content filtering strictness controls safety moderation of potentially harmful content, which is unrelated to whether legal clauses follow a numbering convention. Option D is incorrect because the team has already tried extensive few-shot examples without success, and piling on more examples in every request increases prompt length, cost, and latency while offering diminishing returns on a formatting problem prompting has already failed to solve.