3 What data do I need for model adaptation?

Most organizations have data. Very few have AI-ready data. That distinction matters more than any architecture decision you will make during adaptation, because a model can only learn what its training data teaches it. Feed it inconsistent, unrepresentative records, and you get a model that is confidently wrong in ways that are hard to diagnose after deployment. When adapting a model, having high-quality, representative data is important. Common questions that come up are: How much data do we need? What kind of data do we need to effectively perform model adaptation? Can I use synthetic data to fill data gaps?

3.1 Data readiness spectrum

3.2 Why data quality is important

3.3 How much data do you need?

3.3.1 Base models versus instruct models

3.3.2 How accuracy responds to data volume

3.3.3 Does your training data reflect your production distribution?

3.3.4 Edge cases matter disproportionately

3.3.5 Diversity across user types, conditions, and variations

3.3.6 Diminishing returns: Why the first 100 examples do most of the work

3.4 Data acquisition strategies

3.4.1 Human annotation

3.4.2 Weak supervision

3.4.3 Synthetic data from teacher models

3.4.4 Production log mining

3.5 Data versioning and experiment tracking

3.5.1 Experiment logging

3.5.2 Choosing a tracking tool

3.5.3 Data lineage

3.6 Generating synthetic training data

3.6.1 Step 1: Load seed examples from Hugging Face

3.6.2 Step 2: Build the prompt with style anchoring

3.6.3 Step 3: Calling the teacher model to generate the candidates (one call per category)

3.6.4 Step 4: Quality gate

3.6.5 Step 5: Distribution alignment check

3.6.6 Step 6: Mix with real data and save with a version manifest

3.7 Golden test sets and held-out evaluation

3.7.1 What goes in a golden test set

3.7.2 Building your first golden test set

3.7.3 The red team companion set

3.7.4 Common mistakes

3.8 Removing personal data before training