Fine-Tuning LLMs · Lesson 2
Domain Data Strategy
Fine-tuning quality depends on data quality. Collect real examples of the target task, then clean, deduplicate, remove PII and fix formatting.
Formats: raw domain text for continued pre-training, instruction–response pairs for SFT, and chosen/rejected pairs for preference tuning.
A few thousand high-quality, diverse examples often beat a huge noisy set. Synthetic data from a larger model can help but must be filtered and checked. Always hold out an evaluation set.
Key points
- Quality and diversity beat volume
- Clean, deduplicate and redact PII
- Match data format to the training method
- Hold out an evaluation set