Granite Model Does Not Generate `<|tool_call|>` Format in Response
Description:
Hello,
I am attempting to use the Granite-3.0-8b-Instruct model to generate responses in the <|tool_call|> format for external function calls. While the model successfully generates responses in JSON format, it does not include the <|tool_call|> token.
Steps to Reproduce:
- The following prompt was used:
<|start_of_role|>available_tools<|end_of_role|>
[{'name': 'calculator.add', 'description': 'Adds two numbers.', 'parameters': {'type': 'object', 'properties': {'num1': {'type': 'integer', 'description': 'The first number'}, 'num2': {'type': 'integer', 'description': 'The second number'}}, 'required': ['num1', 'num2']}}]<|end_of_text|>
<|start_of_role|>system<|end_of_role|>The AI must only use the provided tools to respond.<|end_of_text|>
<|start_of_role|>user<|end_of_role|>What is 123 + 456?<|end_of_text|>
<|start_of_role|>assistant<|end_of_role|>
- The modelโs response was:
<|start_of_role|>assistant<|end_of_role|>{
"tool": {
"name": "calculator.add",
"arguments": {
"num1": 123,
"num2": 456
}
}
}<|end_of_text|>
Expected Behavior:
The expected response format was as follows:
<|start_of_role|>assistant<|end_of_role|><|tool_call|>calculator.add({"num1": 123, "num2": 456})<|end_of_text|>
Additional Context:
- The
<|tool_call|>token is defined inspecial_tokens_map.json, so the model should ideally recognize and generate this token. - However, while the model generates valid JSON responses, it does not include the
<|tool_call|>token in its output.
Questions:
- Is there any additional configuration required to make
<|tool_call|>functional with the Granite model? - Could you provide examples of prompt designs that would effectively guide the model to generate the
<|tool_call|>format? - Is the modelโs inability to generate the
<|tool_call|>token a limitation of its training data?
Thank you!
I would appreciate any additional guidance or debugging steps you can provide.
Additional Comment:
Hello,
I wanted to clarify that while the issue was submitted with an English prompt for your convenience, the actual tests were conducted using a Korean prompt. The English prompt in the issue description is a direct translation of the Korean prompt to make it easier for your team to review.
Here is the original Korean prompt I used during testing:
<|start_of_role|>available_tools<|end_of_role|>
[
{
"name": "calculator.add",
"description": "๋ ์ซ์์ ํฉ์ ๊ณ์ฐํฉ๋๋ค.",
"parameters": {
"type": "object",
"properties": {
"num1": {
"type": "integer",
"description": "์ฒซ ๋ฒ์งธ ์ซ์"
},
"num2": {
"type": "integer",
"description": "๋ ๋ฒ์งธ ์ซ์"
}
},
"required": [
"num1",
"num2"
]
}
}
]
<|end_of_text|>
<|start_of_role|>system<|end_of_role|>AI๋ ์ ๊ณต๋ ๋๊ตฌ๋ง ์ฌ์ฉํ์ฌ ์ง๋ฌธ์ ๋ต๋ณํฉ๋๋ค.<|end_of_text|>
<|start_of_role|>user<|end_of_role|>123 + 456์ ์ผ๋ง์ธ๊ฐ์?<|end_of_text|>
<|start_of_role|>assistant<|end_of_role|>
The results I received using this Korean prompt are consistent with the issue described. Thank you for your attention, and please let me know if additional details or clarification are needed!
Additional Comment:
Hello,
I would like to provide additional details regarding the issue. I conducted tests using both English and Korean prompts based on the instructions from the IBM Granite Function Calling tutorial (link).
Here are the results:
English Prompt Test
Prompt:
<|start_of_role|>available_tools<|end_of_role|>
[
{
"name": "get_stock_price",
"description": "Retrieve the current price of a given stock ticker.",
"parameters": {
"type": "object",
"properties": {
"ticker": {
"type": "string",
"description": "The stock ticker symbol (e.g., AAPL for Apple Inc.)."
}
},
"required": ["ticker"]
}
}
]
<|end_of_text|>
<|start_of_role|>system<|end_of_role|>The AI must only use the provided tools to respond.<|end_of_text|>
<|start_of_role|>user<|end_of_role|>What is the current price of IBM stock?<|end_of_text|>
<|start_of_role|>assistant<|end_of_role|>
Response:
<|start_of_role|>assistant<|end_of_role|><|tool_call|>{"name": "get_stock_price", "arguments": {"ticker": "IBM"}}<|end_of_text|>
Korean Prompt Test
Prompt:
<|start_of_role|>available_tools<|end_of_role|>
[
{
"name": "get_stock_price",
"description": "์ฃผ์ด์ง ์ฃผ์ ์ฌ๋ณผ์ ํ์ฌ ๊ฐ๊ฒฉ์ ์กฐํํฉ๋๋ค.",
"parameters": {
"type": "object",
"properties": {
"ticker": {
"type": "string",
"description": "์ฃผ์ ์ฌ๋ณผ (์: AAPL์ Apple Inc.)"
}
},
"required": ["ticker"]
}
}
]
<|end_of_text|>
<|start_of_role|>system<|end_of_role|>AI๋ ์ ๊ณต๋ ๋๊ตฌ๋ง ์ฌ์ฉํ์ฌ ์ง๋ฌธ์ ๋ต๋ณํด์ผ ํฉ๋๋ค.<|end_of_text|>
<|start_of_role|>user<|end_of_role|>IBM์ ํ์ฌ ์ฃผ๊ฐ๋ ์ผ๋ง์ธ๊ฐ์?<|end_of_text|>
<|start_of_role|>assistant<|end_of_role|>
Response:
<|start_of_role|>assistant<|end_of_role|>{
"tool": {
"name": "get_stock_price",
"arguments": {
"ticker": "IBM"
}
}
}<|end_of_text|>
Observations
English Prompt:
- The
<|tool_call|>special token was correctly included in the modelโs response. - The response followed the expected format, indicating proper behavior.
- The
Korean Prompt:
- The
<|tool_call|>special token was not included in the modelโs response. - Instead, the response was generated in a plain JSON format without the special token.
- The
Hypothesis:
- The difference in behavior may be due to the modelโs training data being more optimized for English prompts.
- Korean prompts may not have sufficient training examples for
<|tool_call|>generation.
Questions:
- Does the Granite model currently support
<|tool_call|>generation for non-English prompts, particularly Korean? - If support for non-English prompts is limited, are there plans to improve this functionality in the future?
- Are there any specific adjustments or configurations I can make to ensure
<|tool_call|>is generated for Korean prompts?
Thank you for your assistance! Please let me know if additional details or clarification are needed.