Glossary

Every term has two sentences, an example, and the English word next to the Bulgarian one.

benchmark (тест / бенчмарк)

In two words. The same exam for many models. The score is for the snapshot date and does not say which is best for you.

🧒 For kids: Like a school test: the same questions, different scores.

Where you meet it: Board, archive, model profile.

BSI (BSI)

In two words. The stable article number on bansko.si: year and sequence. It does not change when we correct the text.

🧒 For kids: Like a library book number. It stays the same even if the cover changes.

Where you meet it: Articles, masthead, register.

chatbot (чатбот)

In two words. A window where you type to a model and it replies with text.

🧒 For kids: Like a chat with a friend, except a program is on the other side.

Where you meet it: Task cards, guide.

context (контекст)

In two words. How much of the conversation the model still “holds” in memory. Older text can be dropped.

🧒 For kids: Like a short memory: if the story is long, the start can be lost.

Where you meet it: More details on the profile.

DOI (DOI)

In two words. The permanent “address” of a scientific article. Like a library book number.

🧒 For kids: A special address that does not break if the website changes look.

Where you meet it: Source frame, citation.

Elo (Elo)

In two words. A number from a contest where people compare two answers without knowing which model wrote them. A small gap is often almost a tie.

🧒 For kids: Like chess points: win a match and you go up, lose and you go down.

Where you meet it: Board, LMArena metric.

free limit (безплатен лимит)

In two words. How much you can ask without paying. Limits change; we write them only with a source.

🧒 For kids: Free is not endless. After the limit, asking stops or costs money.

Where you meet it: Guide (price question), model profile.

hallucination (халюцинация)

In two words. When the AI invents something and says it confidently, as if it were true. Example: a made-up book with a real-sounding author.

🧒 For kids: AI sometimes makes things up and still sounds sure. Always check.

Where you meet it: Limits on the recommendation card, Glossary.

model (модел)

In two words. A program trained on a lot of text that predicts the next word. It does not think like a person.

🧒 For kids: Like a well-read parrot: it knows many words, but it still makes mistakes.

Where you meet it: Model profile, guide, test board.

multimodal (мултимодален)

In two words. A model that accepts not only text, but also an image, sound, or file. We write this only with a source.

🧒 For kids: It can look at a picture or listen to sound, not only read words.

Where you meet it: “Pictures” card, Today.

open model (отворен модел)

In two words. A model you can download and run yourself. “Open” does not always mean free for everything.

🧒 For kids: Like a recipe you can cook at home, not only in a restaurant.

Where you meet it: Model profile, archive.

prompt (подкана)

In two words. The text you give the model. A clear prompt usually gets a clearer answer.

🧒 For kids: This is the question or task you tell the AI.

Where you meet it: Task cards, the “How to ask” block.

preprint (препринт)

In two words. A scientific article published before or without review by other scientists. It may contain errors.

🧒 For kids: A draft that scientists release early so others can read it.

Where you meet it: Articles, source frame.

token (токен)

In two words. A small piece of a word that models use to count text. Speed is often “tokens per second”.

🧒 For kids: A token is a little piece of a word. A faster model writes more pieces each second.

Where you meet it: “Which AI is best?” board, model profile.