AI glossary in Bangla

How many examples the model processes together in one training step.

Understanding images and video — detecting objects, counting, reading text (OCR).

How many tokens a model can consider at once — prompt, documents and answer together.

NVIDIA’s platform for running programs on its GPUs — PyTorch and TensorFlow use it under the hood.

An organised collection of examples for training or testing a model.

A model that generates images from text — Stable Diffusion, FLUX.

Turning the meaning of text or an image into a list of numbers, so similar things can be found.

One full pass of the model over the whole training data.

Training an existing model further on your own data so it becomes an expert at your task.

Floating-point calculations per second — a measure of GPU speed.

A processor that does thousands of small calculations at once — the main engine for running and training AI.

When an AI confidently states something false or made up.

Using a trained model to do work — answering a question, generating an image.

Code, results, text and charts on one page — the standard workspace for AI.

An AI trained on huge amounts of text that answers questions, writes and translates — like ChatGPT, Llama or Qwen.

Fine-tuning by training small “adapters” instead of the whole model — less memory, lower cost.

The model memorises its examples and does badly on new ones.

The numbers inside a model that are learned during training. “8B” means 8 billion parameters.

LoRA on top of a 4-bit compressed base model — fine-tuning with the least memory.

Storing parameters in fewer bits to shrink a model — 4-bit needs about a quarter of the memory.

Before answering, the model looks up relevant passages from your documents — your own knowledge, no training needed.

A multi-dimensional table of numbers — all AI data and parameters are stored as tensors.

The small pieces an LLM splits text into. Both price and limits are measured in tokens.

Adjusting a model’s parameters with data until it learns the task.

The design behind almost every modern language and image model — it uses “attention” to relate every part of the input.

Stores embeddings and quickly finds the items closest in meaning.

The GPU’s own memory. If a model does not fit in it, it cannot run.