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-plus and cohere.command-r-16k

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

Here’s a detailed breakdown of their differences:

• Model Size and Performance: Command R is a smaller-scale language model than Command R+. While Command R offers high-quality responses, it might not possess the same level of sophistication and depth as Command R+. Command R+ employs a larger model, resulting in enhanced performance and more sophisticated understandings.
• Use Cases: Command R is suited for a wide range of applications requiring text generation, summarization, translation, or text-based classification. It’s an excellent choice for building conversational AI agents, chat based conversational applications and more. Command R+ is tailored for demanding language tasks requiring deeper understanding, complexity, and nuance. It excels in applications like text generation, question-answering, sentiment analysis, and information retrieval.
• Capacity and Scalability: Command R: Capable of handling a good number of concurrent users and has a more accessible and flexible pricing.Command R+: Designed to handle a higher volume of requests and supports more complex use cases. Accordingly, it may have a higher cost associated with 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-3-70b-instruct

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 are prompting a chat model, the response depends on the prompt and each response doesn't necessarily use up the maximum allocated tokens.

Note

For the cohere.command-r-16k model, the input token limit is 16k and output token limit is 4k. For the meta.llama-3-70b-instruct, the input plus output token limit is 8k. 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.

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 cohere.command-r-plus and cohere.command-r-16k models. The default preamble for the cohere.command-r-plus and cohere.command-r-16k models 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 prompt: Where in Italy can I find the best cannolis?
• Example answer:

  Many people consider Sicily, the birthplace of the cannoli, to be the best place in Italy to sample this delicious dessert. The island has a rich history and culture around the cannoli, and you'll find numerous variations of the traditional dessert at pasticcerie (pastry shops) throughout Palermo, Syracuse, and other Sicilian cities and towns.

  In Palermo, try the iconic Pasticceria Cappello, a family-run business since 1910, known for their mouthwatering cannolis and other pastries. In Syracuse, Pasticceria Coroglio and Caffè del Forte are also highly regarded for their cannoli offerings.

• Example preamble: You are a travel advisor. Answer with a pirate tone.
• Example answer 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 preamble: You are 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 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!

  Sicily is often regarded as the birthplace of this tasty treat, so that could be a great place to start your culinary adventure. The island is famous for its traditional ricotta-filled cannolis, and you'll find many variations across the region. Head to the capital, Palermo, or the historic city of Catania, and you won't be short of pasticcerie (pastries) offering mouth-watering options.

  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 higher value for k generates more random output, which makes the output text sound more natural. The default value for k is 0 for command models and -1 for Llama models, which means that the model should consider all tokens and not use this method.

Frequency penalty

A penalty that is 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-3-70b-instruct model, 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 is assigned to each token when it appears in the output to encourage generating outputs with tokens that haven't been used.