In training the model sees examples, makes a prediction, the error (loss) is measured, and the parameters are nudged to reduce it — millions of times over.
Training a large LLM from scratch takes thousands of GPUs and a fortune, so most people fine-tune an existing model instead — which can take hours and cost a few hundred taka.
Example
Training a classifier on 50,000 Bangla news headlines for a thesis.
বাংলায়: ট্রেনিং (মডেল শেখানো) — ডেটা দেখিয়ে মডেলের প্যারামিটার ঠিক করা, যাতে সে কাজটা শিখে নেয়।