Skip to main content

What happens when I fine-tune a model?

How to choose a model, select datasets, start training, download fine-tuned models, and troubleshoot performance issues.

N
Written by Niko McCarty
Updated this week

In Minibase, training a model means fine-tuning an existing base model with your dataset. Right now, fine-tuning uses a method called LoRA (Low-Rank Adaptation). LoRA is an efficient approach that updates only a small number of parameters instead of retraining the whole model. This makes fine-tuning much faster, cheaper, and more resource-friendly, while still tailoring the model to your dataset.

All models in Minibase—whether task-based, chat-based, language-based, or micro-based—can be fine-tuned using any dataset in the system. The choice of dataset and model combination depends on the kind of behavior you want to achieve.

How to Use

Step 1: Choose your model

  • Go to the Models tab.

  • Select the type of base model you want to fine-tune (Task, Language, Chat, or Micro).

  • Name your model and add an optional description.

Step 2: Select your dataset

  • Upload your dataset via the Datasets tab, or select an existing dataset from your organization.

  • Any dataset you upload into Minibase is compatible with any model type.

  • Choose a dataset that matches the problem you’re trying to solve. For example:

    • Q&A dataset → Task-based model

    • Conversational dataset → Chat-based model

    • Translation dataset → Language-based model

Step 3: Kick off training

  • Start fine-tuning by clicking Train model.

  • Behind the scenes, Minibase applies LoRA fine-tuning to adapt the base model to your dataset.

  • Training time depends on model size and dataset size. Larger models train slower, but offer more capability.

Step 4: Download or use your model

  • Once training is complete, you can:

    • Download the model in a LoRA format compatible with Hugging Face and other frameworks.

    • Use it directly in Minibase to run inference, test in the playground, or integrate into your workflow.

Tips & Best Practices

  • Match dataset to model: Even though any dataset works with any model, results are best when aligned. Use task-style datasets with task models, chat-style with chat models, etc.

  • Use clean, high-quality data: Fine-tuning amplifies your dataset. Typos, inconsistent formatting, or poor examples can lead to unreliable outputs.

  • Start small, then expand: Begin with a focused dataset. Evaluate your model’s behavior, then add more examples to improve coverage.

  • Experiment with naming and versions: Give models descriptive names (e.g., “Customer Support v1”) to keep track of experiments.

Troubleshooting

Training is too slow

Larger models (like Chat-based or Language-based) require more resources. For faster iterations, start with smaller models like Task-based or Micro-based.

My model isn't learning well.

Check dataset quality and size. Around 3,000 examples is a minimum for decent performance; ~10,000 is better for robust results. Make sure your dataset matches the type of task.

Did this answer your question?