Dataset preparation

Data quality checklist for training/fine-tuning

Dataset readiness: 0%

Task defined: what the model should learn
Data collected from permitted sources (right to use)
Personal data removed or anonymized
Format unified: same example structure
Duplicates and garbage removed (dedup)
Class balance: no category <10% unless intended
Validation split set aside (10–20%)
Dataset origin and limitations documented

Examples in dataset

Volume estimate: Enough for fine-tuning

For instruction fine-tuning, 200–2000 quality examples suffice — quality beats quantity. An example should contain: instruction/context → desired response in your style. Golden rule: 100 perfect examples beat 10,000 raw ones.