Understanding parameter settings for supported foundational models
- Dernière mise à jour2024/11/11
Understanding parameter settings for supported foundational models
In all AI models, parameters are numerical values that can be used to configure settings of foundational Models (LLMs) that influence how a model processes information by data analysis and makes predictions via responses. More parameters allow for more complex settings and result in better prompt-responses.
- Amazon Bedrock
- Google Vertex AI
- Azure OpenAI
- OpenAI
| Model | Parameters |
|---|---|
| Amazon Bedrock |
|
| Google Vertex AI |
Remarque : See L'action Vertex AI : Chat AI for parameter details.
|
| Azure OpenAI |
Remarque : See Azure OpenAI: Chat AI for parameter details.
|
| OpenAI |
Remarque : See L'action OpenAI : Chat AI for parameter details.
|
Let us look at the use and functionality of these parameters. Understanding parameter settings helps you assess a model's balance between capability of the model and its performance.
Model parameters explained
Foundational model parameters give you the ability to fine-tune the model's ability to process complex prompt inputs to return more accurate and refined responses. Select a model based on how complex your prompt-input is, so you can configure the parameters accordingly to process and return accurate and well-defined responses.
- Document retrieval count
- This setting for hyperscaler vendors typically refers to the number of documents or data entries that can be retrieved in a single query or operation. This setting is crucial for managing and optimizing the performance of data retrieval processes, especially in large-scale environments like those managed by hyperscalers vendors.
- Frequency Penalty / Presence Penalty
- This setting discourages repetition in the generated text by restricting repetitive use of the tokens based on their frequency of use. The more you use a token in the text, the less likely it will be repeated. Choose a value between -2.0 to 2 with decimal value increments.
- Max Tokens / Max Tokens To Sample / Max Output Tokens
- This setting denotes the maximum number of tokens used in a generated response. You can choose a value between 1 to 2048, with default value set at 2048. Your prompt response will be affected based on the value set here. Allotting more tokes will return a more comprehensive and detailed response.
- Presence Penalty
- This setting discourages repetition of tokens in the generated text by restricting the use of tokens based on how frequently they appear. The more often a token is used in the text, less likely it will be repeated. Choose a value between -2.0 to 2 with decimal value increments.
- Temperature
- A higher value returns diverse and less predictable responses. You can choose a value between 0 to 1 with decimal value increments. This means that with higher value, the responses returned are more varied.
- Top P / Top K
- This setting determines the diversity of the generated response. Higher value returns more diverse responses. We recommend changing either the P/K or Temperature value, not both. Choose a value between 0 to 1 with decimal value increments.
- N
- This defines the number of responses generated by the model for a specific prompt. Choose between 1 to 9 with default value set at 1. If you assign more tokens and configure a higher temperature value and choose a high N value of 9, you will get 9 varied responses with details that would accommodate 2048 tokens.
Return to Créer Compétences en IA.