Fine-tuning in Minibase uses LoRA to adapt base models with your datasets. While the process is straightforward, users sometimes run into questions or confusion during training. This FAQ collects the most common problems and how to solve them.
How to Use
Use this guide whenever you run into an issue during model training.
Each question below expands with troubleshooting steps.
If your problem isn’t listed, you can always reach out to us directly at [email protected] or via Intercom chat.
Troubleshooting
My dataset uploaded fine, but training fails immediately
My dataset uploaded fine, but training fails immediately
This typically means the dataset has a formatting issue. Solution:
Check that your dataset includes valid Instruction, Input, and/or Response fields.
Try removing extra or blank columns.
Make sure file is in CSV, Excel, JSON, or JSON-L format.
I can’t see my trained model in the list
I can’t see my trained model in the list
Sometimes models take a few minutes to appear in your account after training. Solution: Refresh the Models tab. If the model still doesn’t appear after 30 minutes, reach out to [email protected].
Training is taking too long
Training is taking too long
Larger models (like Chat-Based or Language-Based) train slower, especially on large datasets. Solution:
Start with a smaller model (e.g., Micro-Based or Task-Based) to prototype.
Reduce your dataset size when experimenting.
Once satisfied, scale up to bigger models.
My model finished training but the results aren’t good
My model finished training but the results aren’t good
This usually comes down to dataset quality or alignment. Solution:
Make sure your dataset matches your intended use case (e.g., conversational data for the Chat-Based model).
Clean and reformat your examples for consistency.
Add more examples (~3,000 is minimum, ~10,000 is better).
My model training looks stuck at 50% (or another percentage)
My model training looks stuck at 50% (or another percentage)
This does not necessarily mean something is broken. Training on large datasets (e.g., 10,000–50,000 examples) can take several hours to complete. During this time, the progress bar may stay on a single percentage value for a while. Solution:
Be patient and allow the process to finish.
If your training has been idle for more than 12 hours, or if you’re concerned, contact us at [email protected] and we’ll check on your job.