Tips for prompting granite-13b-chat-v2 for conversational tasks

• In follow-up questions, avoid pronouns. For example, ask “What does the variable represent?” instead of “What does it represent?”

• If you want the foundation model to generate a response in table format, explicitly ask in the prompt for the model to return a table that is generated in markdown.

Retrieval-augmented generation prompt example

The following prompt uses the granite-13b-chat-v2 foundation model to answer questions about prompt tuning.

Note: The document content is taken from the Methods for tuning foundation models topic.

<|system|>
You are Granite Chat, an AI language model developed by IBM. You are a cautious assistant. You carefully follow instructions. You are helpful and harmless and you follow ethical guidelines and promote positive behavior.
<|user|>
You are an AI language model designed to function as a specialized Retrieval Augmented Generation (RAG) assistant. When generating responses, prioritize correctness, i.e., ensure that your response is grounded in context and user query. Always make sure that your response is relevant to the question.
Answer length: concise
[Document]
How prompt-tuning works
Foundation models are sensitive to the input that you give them. Your input, or how you prompt the model, can introduce context that the model will use to tailor its generated output. Prompt engineering to find the right prompt often works well. However, it can be time-consuming, error-prone, and its effectiveness can be restricted by the context window length that is allowed by the underlying model.
Prompt-tuning a model in the Tuning Studio applies machine learning to the task of prompt engineering. Instead of adding words to the input itself, prompt-tuning is a method for finding a sequence of values that, when added as a prefix to the input text, improve the model's ability to generate the output you want. This sequence of values is called a prompt vector.
Normally, words in the prompt are vectorized by the model. Vectorization is the process of converting text to tokens, and then to numbers defined by the model's tokenizer to identify the tokens. Lastly, the token IDs are encoded, meaning they are converted into a vector representation, which is the input format that is expected by the embedding layer of the model. Prompt-tuning bypasses the model's text-vectorization process and instead crafts a prompt vector directly. This changeable prompt vector is concatenated to the vectorized input text and the two are passed as one input to the embedding layer of the model. Values from this crafted prompt vector affect the word embedding weights that are set by the model and influence the words that the model chooses to add to the output.
To find the best values for the prompt vector, you run a tuning experiment. You demonstrate the type of output that you want for a corresponding input by providing the model with input and output example pairs in training data. With each training run of the experiment, the generated output is compared to the training data output. Based on what it learns from differences between the two, the experiment adjusts the values in the prompt vector. After many runs through the training data, the model finds the prompt vector that works best.
[End]
How does prompt tuning work?
<|assistant|>

The granite-13b-chat-v2 foundation model generates an accurate summary of how prompt tuning works that is based on information from the document content that was provided with the input.

The output looks something like this:

Prompt tuning is a method for improving a model's ability to generate specific outputs by adjusting a sequence of values called a prompt vector. Instead of directly adding words to the input, prompt tuning crafts a prompt vector that directly influences the word embedding weights set by the model. This is achieved by vectorizing input text and encoding token IDs into a vector representation, which is then concatenated with the crafted prompt vector and passed as one input to the model's embedding layer. The experiment adjusts the values in the prompt vector based on the differences between the generated output and the desired output in training data examples. Over many runs through the training data, the model finds the prompt vector that works best.

Classification

To use the granite-13b-chat-v2 foundation model to classify information, follow these recommendations.

The following table lists the recommended model parameters for prompting the granite-13b-chat-v2 foundation model for a classification task.

To prompt the granite-13b-chat-v2 foundation model for a classification task, try these steps:

1. Identify the classes or classification labels that you want the model to assign to the input. Be sure to list these class labels in the instruction segment of your prompt.

   For example, if you want to classify customer product reviews as positive or negative, you might define two class labels: Postitive and Negative.

2. Collect two or three representative examples of the type of input text that you want the model to classify.

3. Work with the granite-13b-chat-v2 foundation model from the Prompt Lab in freeform mode so that you can structure your prompts.

4. From the Model parameters panel, set the recommended model parameters from Table 3.

5. In your prompt, clearly identify the system prompt, user input, and where the model's output should go.

   For example, the following prompt structure was used when the granite-13b-chat-v2 foundation model was trained to classify text:

   <|system|>
   You are Granite Chat, an AI language model developed by IBM. You are a cautious assistant. You carefully follow instructions. You are helpful and harmless and you follow ethical guidelines and promote positive behavior.
   <|user|>
   {INSTRUCTION}
   Your response should only include the answer. Do not provide any further explanation.
   Here are some examples, complete the last one:
   {INPUT_LABEL}:
   {ICL_INPUT_1}
   {OUTPUT_LABEL}:
   {ICL_OUTPUT_1}
   {INPUT_LABEL}:
   {ICL_INPUT_2}
   {OUTPUT_LABEL}:
   {ICL_OUTPUT_2}
   {INPUT_LABEL}:
   {TEST_INPUT}
   {OUTPUT_LABEL}:
   <|assistant|>
   
   

   You can use a similar structure to leverage the model's training. Simply replace the placeholder variables in the prompt template.

   Table 3a: Classification template placeholder variables

   Placeholder variable	Description	Examples
   {INSTRUCTION}	Description of the task. Include a list of the classes that you want the model to assign to the input.	For each product review, indicate whether the review is Positive or Negative.
   {INPUT_LABEL}	Short label for the text to be classified.	Input, Customer review, Feedback, Comment
   {OUTPUT_LABEL}	Short label that represents the classification value.	Class
   {ICL_INPUT_N}	Optional. Examples of input text to be classified. Add examples when you want to use a few-shot prompt to support in-context learning.	The service representative did not listen to a word I said. It was a waste of my time.
   {ICL_OUTPUT_N}	Example outputs with class labels assigned to the corresponding input text examples.	Positive, Negative

<|system|>
You are Granite Chat, an AI language model developed by IBM. You are a cautious assistant. You carefully follow instructions. You are helpful and harmless and you follow ethical guidelines and promote positive behavior.
<|user|>
For each feedback, specify whether the content is Positive or Negative. Your response should only include the answer. Do not provide any further explanation.
Here are some examples, complete the last one:
Feedback:
Carol, the service rep was so helpful. She answered all of my questions and explained things beautifully.
Class:
Positive
   
Feedback:
The service representative did not listen to a word I said. It was a waste of my time.
Class:
Negative
   
Feedback:
Carlo was so helpful and pleasant. He was able to solve a problem that I've been having with my software for weeks now.
Class:
<|assistant|>

Extraction

To use the granite-13b-chat-v2 foundation model to extract information, follow these recommendations.

The following table lists the recommended model parameters for prompting the granite-13b-chat-v2 foundation model for an extraction task.

To prompt the granite-13b-chat-v2 foundation model for an extraction task, try these steps:

1. Identify the information types that you want the model to extract from the input. Be sure to list these information type labels in the instruction segment of your prompt.

   For example, if you want to extract key pieces of information from a company's US Securities and Exchange Commission 10-K form, you might identify an information type such as a Line Of Credit Facility Maximum Borrowing Capacity.

2. Collect 2 or 3 representative examples of input text with the type of information that you want the model to extract.

3. Work with the granite-13b-chat-v2 foundation model from the Prompt Lab in freeform mode so that you can structure your prompts.

4. From the Model parameters panel, set the recommended model parameters from Table 4.

5. Clearly identify the system prompt, user input, and where the model's output should go.

   For example, the following prompt structure was used when the granite-13b-chat-v2 foundation model was trained to extract information from text:

   <|system|>
   You are Granite Chat, an AI language model developed by IBM. You are a cautious assistant. You carefully follow instructions. You are helpful and harmless and you follow ethical guidelines and promote positive behavior.
   <|user|>
   {INSTRUCTION}
   Your response should only include the answer. Do not provide any further explanation.
   Here are some examples, complete the last one:
   {INPUT_LABEL}:
   {ICL_INPUT_1}
   {OUTPUT_LABEL}:
   {ICL_OUTPUT_1}
   {INPUT_LABEL}:
   {ICL_INPUT_2}
   {OUTPUT_LABEL}:
   {ICL_OUTPUT_2}
   {INPUT_LABEL}:
   {TEST_INPUT}
   {OUTPUT_LABEL}:
   <|assistant|>
   
   

   You can use a similar structure to leverage the model's training. Simply replace the placeholder variables in the prompt template.

   Table 4a: Extraction template placeholder variables

   Placeholder variable	Description
   {INSTRUCTION}	Description of the task. Include a list of the information types that you want the model to extract from the input.
   {INPUT_LABEL}	Short label for the text to be classified.
   {OUTPUT_LABEL}	Short label that represents the extracted value.
   {ICL_INPUT_N}	Optional. Examples of input text with information types to be extracted. Add examples when you want to use a few-shot prompt to support in-context learning.
   {ICL_OUTPUT_N}	Example outputs with information types extracted from the corresponding inputs.

<|system|>
You are Granite Chat, an AI language model developed by IBM. You are a cautious assistant. You carefully follow instructions. You are helpful and harmless and you follow ethical guidelines and promote positive behavior.
<|user|>
Extract the Line Of Credit Facility Maximum Borrowing Capacity from the 10K sentences.
Your response should only include the answer. Do not provide any further explanation.
Here are some examples, complete the last one:
10K Sentence:
The credit agreement also provides that up to $500 million in commitments may be used for letters of credit.
Line Of Credit Facility Maximum Borrowing Capacity:
$500M

10K Sentence:
In March 2020, we upsized the Credit Agreement by $100 million, which matures July 2023, to $2.525 billion.
Line Of Credit Facility Maximum Borrowing Capacity:
$2.525B

10K Sentence:
We prepared our impairment test as of October 1, 2022 and determined that the fair values of each of our reporting units exceeded net book value by more than 50%. Among our reporting units, the narrowest difference between the calculated fair value and net book value was in our Principal Markets segment's Canada reporting unit, whose calculated fair value exceeded its net book value by 53%. Future developments related to macroeconomic factors, including increases to the discount rate used, or changes to other inputs and assumptions, including revenue growth, could reduce the fair value of this and/or other reporting units and lead to impairment. There were no goodwill impairment losses recorded for the nine months ended December 31, 2022. Cumulatively, the Company has recorded $469 million in goodwill impairment charges within its former EMEA ($293 million) and current United States ($176 million) reporting units. Revolving Credit Agreement In October 2021, we entered into a $3.15 billion multi-currency revolving credit agreement (the "Revolving Credit Agreement") for our future liquidity needs. The Revolving Credit Agreement expires, unless extended, in October 2026. Interest rates on borrowings under the Revolving Credit Agreement are based on prevailing market interest rates, plus a margin, as further described in the Revolving Credit Agreement. The total expense recorded by the Company for the Revolving Credit Agreement was not material in any of the periods presented. We may voluntarily prepay borrowings under the Revolving Credit Agreement without premium or penalty, subject to customary "breakage" costs. The Revolving Credit Agreement includes certain customary mandatory prepayment provisions. Interest on Debt Interest expense for the three and nine months ended December 31, 2022 was $27 million and $65 million, compared to $18 million and $50 million for the three and nine months ended December 31, 2021. Most of the interest for the pre-Separation period presented in the historical Consolidated Income Statement reflects the allocation of interest expense associated with debt issued by IBM from which a portion of the proceeds benefited Kyndryl.
Line Of Credit Facility Maximum Borrowing Capacity:
<|assistant|>