Photo by Author | Ideogram
Introduction
Large models of tongue In recent years, the entire scenario of artificial intelligence has revolutionized, which has identified the beginning of a new era in the history of AI. Usually referred to by their acronym LLMS, they changed the way they interact with the machines, whether to retrieve information, ask questions, or produce a variety of human language content.
Since LLM makes our daily and professional life faster, it is very important to understand the concepts and foundations around them, they are both in terms of construction and practical use and applications.
In this article, we discover the terms of the 10 major language model, which are the key to understanding these strong AI systems.
1. Transformer architecture
Applause: Transformer is the basis of large language models. It is a deep nerve network architecture that is picked up for its advanced, which includes numerous components and layers, such as the position, according to the position, allowing the efficient processing of the input settings and familiar with the context of the input settings.
Why is this key?: Thanks to transformer architecture, it has been possible to understand complex inputs of the language and develop an output of language at an extraordinary level, which has to overcome the limits of the former modern natural language processing solution.
2. The focus procedure
Applause: In fact, it is often considered for language translation works in nerve networks, the focus method analyzes the compatibility of each element, which in another setting contains both different lengths and complexity in one sequence. Although the basic focus method is usually not part of LLM’s primary transformer architecture, they laid the foundation for a better view (as we will talk soon).
Why is this key?: Source and target text streams in tasks such as translation and abstract mechanisms are key to aligning text streams, transforming language understanding and breeding process into extreme context.
3. Self -made
Applause: If the transformer is a kind of component within the architecture that is primarily responsible for the success of the LLM, it is a self -made procedure. The self-made traditional focus is overcome the limits of the method, such as each word-token token, more obviously in all other words (tokens), regardless of their status, in relation to all other words (token).
Why is this key?: Focusing on the dependence, samples and mutual relationships between the elements of the same continuity, the deeper meaning of understanding the input layout is incredibly useful for extracting the context, as well as the target setting is being created as a reaction-thus making it more integrated and well-known.
4. Encoder and decoader
Applause: Classic transformer architecture is divided into almost two major components or half: encoder and decoder. The encoder is responsible for processing and encoded the input sequence, which is focused on producing the output layout using the two parts of the encoder output and the results of the encoder, represented by the encoder output. The two parts are connected to each other, so that the decoder receives the results of the input from the encoder (called hidden states). In addition, both encoder and decoder are “copied” in the form of multiple encoder layers and decoder layers respectively: this level of depth helps to learn more abstract and proportional properties of the input and output layout.
Why is this key?: An encoder and a combination of decoder, each with his own components, is the key to balance the input understanding with output generation in LLM.
5. Pre -Training
ApplauseJust as the foundations of a home from the beginning, pre -training is the first time LLM training, that is, gradually learning all its model parameters or weights. The breadth of these models is such that they can take billions of parameters. Therefore, pre -training is a naturally expensive process that takes weeks to weeks to complete and the text data requires a large and diverse corpora.
Why is this key?: Pre -training is very important for the construction of LLM, which can be understood and understood by ordinary language patterns and terminology, which is in the wider field of topics.
6. Fine toning
ApplauseUnlike pre -training, fine toning is the process of taking pre -trained LLMs and re -training on specific sets related to small and more domain, thus specializing in a particular domain or work. Although computational is still expensive, it is less expensive than the pre -training model, and it often has to update the weight of the model in specific layers of architecture, rather than updating the entire set of the entire parameters of the model.
Why is this key?: LLM must specialize in very concrete tasks and domains such as legal analysis, medical diagnosis, or customer support, as the pre -trained model can be reduced to accuracy, terms and compliance requirements.
7. Settlement
Applause: Machines and AI models do not truly understand the language, but are just numbers. It also applies to LLM, so when we usually talk about models that “understand and produce the language”, what they do, they handle the numerical representation of a language that maintains its key features: this numerical (vector, more precise) is what we say.
Why is this key?: Ambing representations enable the input text stream maping LLM to potentially argue, analysis of matching, and data, all of them without losing the main features of the original text. Therefore, the raw reaction created by the model can be mapped in a cement and suitable human language.
8. Instant engineering
Applause: LLM closing users should be familiar with the best ways of using these models to achieve their goals, and quick engineering stands as a strategic and practical approach for this purpose. Instant engineering has added a set of guidelines and techniques to design an effective user indicators that guide the model to the preparation of a useful, accurate and purpose -based response.
Why is this key?: Often, high quality, precise, and obtaining related LLM output is largely a matter of learning that it is a method of writing high quality indicators that are clear, specific and configured to align the capabilities and powers of LLM, such as converting a vague user question in precise and meaningful answer.
9. Learning context
Applause: It is also called some shot learning, it is a way of teaching LLMS to perform new tasks that predict the model, without re -training or fine toning, providing direct indications of the desired results and instructions. It can be considered as a special form of engineering immediately, as it fully benefits the knowledge of the model during pre -training to remove the samples and adapt to the new tasks.
Why is this key?: Flex has been proven to be an effective approach to learning in a flexible and effective way to solve new tasks based on examples of learning in context.
10. Parameter counting
Applause: The size and complexity of the LLM is usually measured by several factors, the parameter count is one of them. Leading model names such as GPT -3 (with 175B parameters) and Lama 2 (with 70B parameters) reflect the importance and importance of the number of parameters in scaling and clearly show the importance and importance of LLM expression in language -producing language. The number of parameters is important when it comes to measuring LLM capabilities, but other aspects such as training data, architecture design, and other aspects of fine toning points are similar.
Why is this key?: The parameter count is important not only in the model “store” and handling linguistic knowledge, but also in challenging reasoning and its performance on breeding tasks, especially when they include multi -phase dialogues between the user and the model.
Wrap
This article detects the importance of ten important terms around large language models: Due to the remarkable achievements made by these models in the past few years, the focus of the entire AI landscape. Being familiar with these ideas, you are in a beneficial position to stay close to new trends and advancements in the landscape of LLM land, rapidly developed.
Ivan Palomars Carcosa AI, Machine Learning, Deep Learning and LLMS is a leader, writer, speaker, and adviser. He trains and guides others to use AI in the real world.