Skip to main content

Advanced LLM Settings

LLM Personalization Options​

Features We Recommend Using When Customizing Your Own Language Model in SkyStudio

Large Language Models (LLMs) are tools that have revolutionized the field of natural language processing (NLP). These models deliver high performance on many tasks such as text generation, translation, and text completion. When using an LLM, it is necessary to understand and correctly configure certain important settings in order to optimize the model's behavior and performance. SkyStudio was developed by Skymod Technology so that you can carry out this task in the easiest possible way. In this article, we will explain the "Balanced," "Precise," and "Creative" options you will encounter when selecting an LLM in SkyStudio, along with the settings associated with these options.

LLM Personalization Options: Balanced, Precise, and Creative

Balanced​

The Balanced option makes the selected model provide responses in a balanced way. This setting is generally suitable for general-purpose use and enables the model to produce both accurate and creative responses. Balanced settings are designed to perform well in a variety of scenarios.

Precise​

The Precise option focuses on making the selected model provide more accurate and precise responses. This setting is preferred especially in situations that require technical or sensitive information. It is ideal for users who want the model's responses to be accurate and reliable. In particular, for assistants that are meant to specialize in a specific subject and are trained with the relevant domain documents, it can prevent the model from going out of scope.

Creative​

The Creative option makes the selected model provide more creative and original responses. This setting is used for tasks such as creative writing, storytelling, or generating innovative ideas. By offering more variety and originality, the model contributes to creative processes.

Parameters for Determining the Model's Behavior and the Characteristics of Its Responses​

There are some fundamental settings for determining the model's behavior and the characteristics of its responses. Here are the most important of these settings:

Creativity (Temperature): Temperature controls the randomness of the model's responses. A low temperature value causes the model to give more consistent and predictable responses, while a high temperature value produces more varied and creative responses. For example, when the temperature value is 0.2, the responses are more precise and consistent, whereas when it is 0.8, the responses can be more creative and varied.

Response Length (Max Tokens): Max tokens determines the maximum number of tokens (words or symbols) the model can produce. This setting is used to control the length of the responses. For example, when the max tokens value is set to 100, the model can produce responses of at most 100 tokens in length. It helps.

Top P: Top-p controls the size of the token pool the model uses to select the most probable responses. This setting helps the model produce more reliable and consistent responses. A low top-p value makes the model select only the most probable tokens.

Presence Penalty: The presence penalty makes responses more varied by penalizing the presence of a particular token. This setting is used especially to prevent certain words from being overused.

Frequency Penalty: The frequency penalty reduces the likelihood of the model repeating the same token. This setting makes responses more varied and helps avoid repetition.