How accurate is this estimate?

It is a planning heuristic, not a guarantee. Real requirements depend on data quality, task difficulty, and label noise. Use the number to size your effort, then validate with a small pilot run and a held-out eval set.

Why does task type matter so much?

Teaching a model a fixed output format or a consistent tone needs far fewer examples than teaching it new domain knowledge or fine-grained classification. The estimator weights each task type accordingly.

What does the baseline slider change?

If the base model is already decent at your task, you mainly need examples to nudge its behaviour, so the count drops. If it is poor, you need more examples to move the needle, so the count rises.

Should I collect the recommended amount before starting?

No. Start near the minimum, measure on a held-out set, and add data only if the eval shows you have not reached your target. Over-collecting up front wastes labelling budget.

Is more data always better?

Not if it is noisy or redundant. A few hundred clean, diverse, correctly labelled examples usually beat thousands of low-quality ones. Prioritise quality and coverage over raw volume.

Fine-Tuning Dataset Size Estimator

Fine-tuning dataset size estimator

“How many examples do I need?” is the first question of every fine-tuning project, and the honest answer is “it depends.” This tool turns that into a concrete starting number by weighing the three factors that matter most: what kind of task it is, how good the base model already is, and how high you need accuracy to go.

How it works

Each task type carries a base example count reflecting its difficulty — teaching a fixed output format or a writing style needs far fewer examples than teaching new classification boundaries or domain knowledge. The estimator then scales that figure by your baseline capability (a strong base model needs fewer examples to nudge) and by your target accuracy (chasing the last few percent costs disproportionately more data). The result is a minimum to start with and a recommended ceiling to plan toward.

Tips and notes

Start at the minimum, not the maximum. Collect the smaller number, run a pilot fine-tune, and measure on a held-out set before labelling more.
Quality beats quantity. A few hundred clean, diverse examples usually outperform thousands of noisy ones. Deduplicate and balance your classes.
Match the eval to the goal. Your held-out set should reflect real production inputs, or your accuracy number will mislead you.
Reserve a test split. Always hold back data the model never sees during training so your reported accuracy is honest.