Skip to main content
What is Fine Tuning? Fine Tuning is a way to teach an AI model using your own data. In CallKaro, this data is usually the transcripts of recorded calls. Why do we Fine Tune? We fine tune to make the AI respond the way we want, not in a generic way. Example: an agent handles a call really well and the customer converts. You want the AI to always respond like that. So you take that call, train the model on it, and the AI learns that style. A well fine-tuned model always knows what to say, which leads to more conversions.

​
How to Fine Tune — Step by Step

  1. Go to Call History. Open the dashboard and go to the “Call History” section.
  2. Pick the right call. Select the call recording you want the model to learn from. Make sure it is a good, ideal call as per your agent’s setup.
  3. Click “Fine Tune LLM”. You will see a button called “Fine Tune LLM” — click it.
  4. Edit the transcript (optional). You can manually edit the call transcript here. Fix it up to make it a perfect training example for the model.
  5. Add to training data. Scroll down the transcript and click “Add to Training Data”. You get two options: Bucket (choose an existing group of training data to add this call into) or Create a New Bucket (make a brand new group). Think of a Bucket like a folder that holds all the calls you want to train the model on.
  6. Go to the “Fine Tuning” section. Open the sidebar and click on “Fine Tuning”.
The Fine Tuning section has these main parts:
  • Fine Tune Jobs: shows the history of all previous fine tuning jobs you have run.
  • Filters: use these to narrow down which calls to train on — Date, Bucket, Call Type, Agent, Conversion Status, Disconnect Reason, and Call Duration. You can also Search, Reset Filters, or Export the data as JSON.
  • Fine Tune Panel: where the actual training happens.
In the Fine Tune Panel you get the following options:
  • Base Model: the AI model you want to fine tune.
  • Fine-Tuned Model Name: give a name to your newly trained model.
  • Learning Rate: how fast the model learns from the data. Too fast and it misses details; too slow and it takes forever. Always aim for the sweet spot, known as the Optimal Learning Rate.
  • Batch Size: how many calls are sent to the model at once during training. Number of Batches = Total Data divided by Batch Size.
  • Epochs: how many times the model goes through all the training data. More epochs means higher cost.