About the Chat Models in Generative AI

Select a model for your chat conversation based on the model size, your project goal, cost, and the style of the model's response. Use the playground's provided examples with each listed model to get a feel for how each model responds to the same prompt and then decide which model's response style goes well with your use case.

The following pretrained foundational models are available in OCI Generative AI for chat:

• cohere.command-r-16k
• cohere.command-r-plus
• meta.llama-3.1-70b-instruct
• meta.llama-3.1-405b-instruct
• meta.llama-3-70b-instruct

Here are some guidelines to help you select the best model for your use case.

Cohere Command R Family

The cohere.command-r-plus (Command R+) and cohere.command-r-16k (Command R) chat models both belong to the Cohere Command R family of LLMs, but they differ in their capabilities, use cases, and pricing. Key differences include:

• Model Size and Performance: Command R is a smaller-scale language model than Command R+. While Command R offers high-quality responses, the responses might not have the same level of sophistication and depth as the Command R+ responses. Command R+ is a larger model, resulting in enhanced performance and more sophisticated understandings.
• Use Cases: Command R is suited for various applications, including text generation, summarization, translation, and text-based classification. It's an ideal choice for building conversational AI agents and chat-based applications. Command R+, on the other hand, is designed for more complex language tasks that require deeper understanding and nuance, such as text generation, question-answering, sentiment analysis, and information retrieval.
• Capacity and Scalability: Command R can handle a moderate number of concurrent users compared to Command R+. Command R+, however, is designed to handle a higher volume of requests and support more complex use cases, which might result in higher prices because of its increased capacity and performance.

In summary, Command R is an excellent choice for those looking for a more affordable and flexible option for general language tasks. On the other hand, Command R+ is designed for power users who require advanced language understanding, higher capacity, and more nuanced responses. The choice between the two would depend on the specific requirements and budget of your application.

Meta Llama 3.1 Family

The meta.llama-3.1-405b-instruct and meta.llama-3.1-70b-instruct belong to the Meta Llama 3.1 family of LLMs with the following key features:

• Model Sizes: 405 and 70 billion parameters
• Context Length: 128,000 tokens, which is 16 times increase from the Meta Llama 3 models
• Multilingual Support: English, French, German, Hindi, Italian, Portuguese, Spanish, and Thai

meta.llama-3.1-405b-instruct

This 405 billion-parameter model is a high-performance option that offers speed and scalability. Compared to the meta.llama-3.1-70b-instruct model, it can handle a higher volume of requests and support more complex use cases. Key features of this model include:

• Recognized as the largest publicly available large language model at the time of its release.
• Suited for enterprise-level applications and research and development initiatives.
• Shows exceptional capabilities in areas such as general knowledge, synthetic data generation, advanced reasoning, and contextual understanding, and long-form text, multilingual translation, coding, math, and tool use.

meta.llama-3.1-70b-instruct

This 70 billion-parameter generation model is perfect for content creation, conversational AI, and enterprise applications. Its key strengths include:

• Summarizing, rewording, and classifying text with high accuracy
• Sentiment analysis and language modeling capabilities
• Effective dialogue systems
• Code generation

meta.llama-3-70b-instruct (deprecated)

This 70 billion-parameter generation model is highly performant and has a broad general knowledge, making it suitable for various tasks, from generating ideas to refining text analysis and drafting written content, such as emails, blog posts, and descriptions.

When using the chat models, you can vary the output by changing the following parameters.

Maximum output tokens

The maximum number of tokens that you want the model to generate for each response. Estimate four characters per token. Because you're prompting a chat model, the response depends on the prompt and each response doesn't necessarily use up the maximum allocated tokens. For each model's maximum output tokens, see the key features in Pretrained Foundational Models in Generative AI.

Note

For the cohere.command-r-16k model, the input token limit is 16,000 and output token limit is 4,000. For the meta.llama-3-70b-instruct model, the input plus output token limit is 8,000. If you don't set a value for the Maximum output tokens parameter, the conversation might use the model's full context length which is the input plus output token limit. For the remaining offered Meta Llama models, the context length for input plus output is 128,000 tokens.

Preamble override

An initial context or guiding message for a chat model. When you don't give a preamble to a chat model, the default preamble for that model is used. You can assign a preamble in the Preamble override parameter, for the models. The default preamble for the Cohere Command R family is:

You are Command.
You are an extremely capable large language model built by Cohere. 
You are given instructions programmatically via an API
that you follow to the best of your ability.

Overriding the default preamble is optional. When specified, the preamble override replaces the default Cohere preamble. When adding a preamble, for best results, give the model context, instructions, and a conversation style.

Tip

For chat models without the preamble override parameter, you can include a preamble in the chat conversation and directly ask the model to answer in a certain way. For example, "Answer the following question in a pirate tone. Where's the best place to go sailing?"

Example 1

• Example preamble: You're a travel advisor. Answer with a pirate tone.
• Example prompt: Where can I find the best cannolis in Italy?
• Example response using the preceding preamble: Argh, me heartie! Ye be wantin' some of the finest cannolis in the land, do ye? Ye should set sail for the lush lands of Sicily, where the streets be lined with delicious pastries and the smells of sweet ricotta fill the air. In the port town of Palermo, ye can find some 'o the finest cannoli kitchens. The locals be claims that the best cannolis be found at Pastificio Graffeo, where the shells be crispy and filled with creamy ricotta.

Example 2

• Example preamble: You're a seasoned marketing professional with a deep understanding of consumer behavior and market trends. Answer with a friendly and informative tone, sharing industry insights and best practices.
• Example prompt: Where can I find the best cannolis in Italy?
• Example response to the preceding preamble: Cannolis are a delicious Italian dessert, and finding the best ones really depends on your personal preferences and what makes the perfect cannoli for you! In fact, Sicily boasts some renowned cannoli hotspots. Caffe Sierra in Palermo has a reputation for exceptional cannolis, as does Pasticceria Cappello, which has multiple branches across the city. These spots are known for their traditional recipes and high-quality ingredients, using fresh ricotta cheese and thin, crispy pastry shells.

Temperature

The level of randomness used to generate the output text.

Tip

Start with the temperature set to 0 or less than one, and increase the temperature as you regenerate the prompts for a more creative output. High temperatures can introduce hallucinations and factually incorrect information.

Top p

A sampling method that controls the cumulative probability of the top tokens to consider for the next token. Assign p a decimal number between 0 and 1 for the probability. For example, enter 0.75 for the top 75 percent to be considered. Set p to 1 to consider all tokens.

Top k

A sampling method in which the model chooses the next token randomly from the top k most likely tokens. A high value for k generates more random output, which makes the output text sound more natural. The default value for k is 0 for Cohere Command models and -1 for Meta Llama models, which means that the model should consider all tokens and not use this method.

Frequency penalty

A penalty that's assigned to a token when that token appears frequently. High penalties encourage fewer repeated tokens and produce a more random output.

For the Meta Llama family models, this penalty can be positive or negative. Positive numbers encourage the model to use new tokens and negative numbers encourage the model to repeat the tokens. Set to 0 to disable.

Presence penalty

A penalty that's assigned to each token when it appears in the output to encourage generating outputs with tokens that haven't been used.