step 03

Training

Train a language model on your dataset. Kanha handles the GPU infrastructure. You just pick a size and click train.

Model sizes

Choose a model size based on your needs and your users’ devices. Larger models can handle more complex material but use more local resources.

Size	Browser memory check	What it is for
Small	~2 GB	Fastest to load and lightest on the device. Best for focused Q&A and FAQ bots.
Medium	~4 GB	More room for nuanced questions, at higher device demands.
Large	~8 GB	The most detailed answers of the three, for the most capable devices.

The SDK applies these approximate thresholds only when the browser reports device memory. Runtime behavior and performance also depend on the local runtime and device.

Model selection

Each trained size is a separate model. The dashboard provides one install snippet for its selected active model and Playground links for other active models. The SDK does not select a size based on device RAM or automatically switch models.

Start a training job

01Open your bot

Go to your bot in Dashboard → Bots. Your bot needs to be linked to a site that has a completed dataset. See the previous step.

02Select dataset and sizes

Select the dataset and the model sizes you want to train. You can select multiple sizes at once. Each creates a separate training job. All sizes are available on every plan.

03Start training

Click Start Training. All jobs are queued and processed sequentially on GPU infrastructure. Your bot status changes to Training immediately.

If the run would go past the training your plan includes, Kanha shows you the extra cost first and waits for you to confirm before anything starts. Nothing is billed until you accept.

What happens during training

Behind the scenes, Kanha runs a training pipeline on your dataset.

01DownloadYour JSONL dataset is fetched to the GPU worker.
02Fine-tuneYour model is trained on your data.
03OptimizeThe model is optimized for on-device execution.
04UploadModel artifacts are uploaded to CDN for fast delivery.

Training usually finishes in 15 to 40 minutes depending on dataset size and model. You’ll see the status update in real time on the bot page, including whether the run is still waiting for capacity or is actively training, plus how long it has been going.

You can cancel a run at any time from the bot page while it is queued or in progress. Cancelling frees the slot straight away and leaves any previously trained models untouched.

When training completes

If you trained multiple sizes, your bot becomes Active as soon as at least one training job succeeds. Each completed job produces a separate trained model. You can test active models from their Playground links. Public installation follows the single generated snippet for the dashboard’s selected active model.

The trained model can run on the visitor’s device through a compatible local runtime.

To retrain with updated content, generate a new dataset from your site’s latest pages and start a new training job. Previous models remain available until replaced. Each training job counts against your monthly train quota (Starter: 3/mo, Pro: 8/mo, Business: 20/mo). If you expect to need more, contact sales to confirm capacity and pricing.