​
1. What is fine-tuning?

Fine-tuning is the process of adapting a pre-trained large language model (LLM) to perform more specialized tasks by training it further on task-specific data. This helps the model become more accurate and efficient at tasks in particular domains or applications.

​
2. When should I fine-tune a model?

You should consider fine-tuning a model when:
  • You need more accuracy on specific tasks or within a specialized domain (e.g., legal, medical, or technical industries).
  • The existing pre-trained model doesn’t perform well on your specific use case, such as company-specific tasks, low-resource languages, or niche topics.
  • You want to improve the model’s performance on tasks with limited available data (low-resource applications).

​
3. What are the advantages of fine-tuning?

Fine-tuning offers several benefits, such as:
  • Improved accuracy for specific tasks or domains.
  • Reduced token usage by minimizing the need for complex prompts.
  • Faster response times and lower costs due to optimized prompt engineering.
  • Better handling of specialized tasks like language localization, domain-specific knowledge integration, or adapting to few-shot learning.

​
4. What models can I fine-tune with Lumino?

Lumino supports fine-tuning for Llama 3.1 models in various sizes, including:
  • Llama 3.1 8B
  • Llama 3.1 70B

​
5. How do I prepare my dataset for fine-tuning?

To prepare your dataset for fine-tuning, ensure it reflects the scenarios your model will handle. You can use a structured format like conversations or interaction examples, where each message has a clearly defined role (e.g., user or assistant). Pay attention to edge cases where the model may have previously underperformed and include those in the dataset.

​
6. How do I upload my dataset to Lumino?

You can upload your dataset through the Lumino Dashboard by navigating to the “Datasets” tab and clicking “Upload Dataset.” Follow these steps:
  1. Enter a name for your dataset.
  2. Add an optional description.
  3. Upload the dataset file. Once uploaded, you’re ready to fine-tune the model.

​
7. How long does fine-tuning take?

The time required for fine-tuning depends on several factors, including the size of the model, the dataset’s size, and the complexity of the task. Smaller models or datasets typically take less time, while larger, more complex tasks may require more computation.

​
8. How do I know if fine-tuning is successful?

After fine-tuning, you can evaluate the model’s performance by running it on a test set. Compare its outputs against the baseline performance to assess improvements. Iterate on the training process if necessary, adjusting the dataset or training parameters to improve results.

​
9. Can I fine-tune multiple models at once?

Yes, with Lumino, you can fine-tune multiple models, such as different sizes of Llama 3.1, on separate tasks or datasets. You can also run evaluations on each to determine which model performs best for your specific needs.

​
10. What are the costs involved in fine-tuning?

The costs of fine-tuning can vary depending on factors such as the model size, dataset size, and computational resources required. Lumino offers detailed pricing for fine-tuning tasks on our Pricing page. Fine-tuning smaller models or optimizing tasks for efficiency can help reduce costs.