​
🧠 Why This Matters

Good models start with good data. Whether you’re fine-tuning, evaluating, or experimenting, following a few simple best practices can make a big difference. This guide helps you:
  • ✅ Avoid common mistakes in dataset preparation.
  • 📊 Set up reliable validation processes.
  • 🔍 Understand how much data is enough.
  • 🛠️ Use your synthetic data effectively — without biasing your results.
These are the foundations you’ll want to get right before you train or evaluate anything in Prem Studio.

​
✅ Key Best Practices

1

Always Have a Validation Dataset

A validation dataset is essential if you want to measure model quality.
  • Use it during fine-tuning to quickly monitor the model performance.
  • Use it to benchmark different models with our agentic evaluator.
If you’re not using a validation dataset, you won’t be able to run evaluations in Prem Studio.
2

Avoid Using Synthetic Data for Validation

Synthetic data is great for training, but avoid using it for validation:
  • It can inflate performance metrics.
  • It might not capture the full complexity of real-world inputs.
Only use high-quality, real data in your validation set to ensure reliable results.
3

How Much Data Do You Actually Need?

The amount of data you need depends on the complexity of your task. Text classification or simple QA tasks may require fewer examples, while nuanced or open-ended tasks typically need more.As a rule of thumb:
  • Start with a few hundred training examples to test your setup.
  • Start with a few dozens validation examples.
Having too few validation examples can lead to noisy evaluation results — both during fine-tuning (unreliable loss function) and during agentic evaluation, where a small sample may be insufficient to draw meaningful conclusions.
After initial fine-tuning, you can assess whether more data is needed by:
  • Reviewing the training loss curve.
  • Running evaluations on your validation set.
  • Checking performance with agentic evaluation tools in Prem Studio.
If performance plateaus early or generalization is poor, consider expanding your dataset with more examples or better coverage. If you don’t have more data, you can use enrich your dataset using synthetic data augmentation to increase the size of your dataset.
4

Split Before You Augment

GIF showing autosplit before data augmentationDon’t augment your dataset and then split it — that leads to data leakage.
  • First, split your dataset into training and validation sets.
  • Then apply synthetic augmentation only to the training split.
If you augment before splitting, you risk data leakage — original datapoints used to generate synthetic examples may end up in your validation set. This compromises evaluation by leaking information from training into validation, leading to inflated and misleading metrics.

​
📦 What’s Next?

Now that your dataset follows these practices, you’re ready to:

​
💡 Pro Tips

  • Name your dataset clearly to reflect its purpose or use case.
  • Use autosplit for quick, clean train/validaiton separation.
  • Periodically re-check your validation set to make sure it reflects real-world data.