tool-call: Qwen 2.5 Instruct also requires object arguments

ggerganov · Sep 28, 2024 · bc3e0c0 · bc3e0c0
1 parent b10ef04
commit bc3e0c0
Show file tree

Hide file tree

Showing 5 changed files with 11 additions and 11 deletions.
diff --git a/common/chat-template.cpp b/common/chat-template.cpp
@@ -36,7 +36,7 @@ llama_chat_template::llama_chat_template(const std::string & chat_template, cons
     _supports_tools = chat_template.find("tools") != std::string::npos;
     _requires_object_arguments =
         chat_template.find("tool_call.arguments | items") != std::string::npos
-        || chat_template.find("{{- tool_call.arguments | tojson }}") != std::string::npos;
+        || chat_template.find("tool_call.arguments | tojson") != std::string::npos;
     _supports_system_role = chat_template.find("System role not supported") == std::string::npos;
 
     if (chat_template.find("<tool_call>") != std::string::npos) {

diff --git a/tests/chat/goldens/Qwen-Qwen2.5-7B-Instruct-tool_use.txt b/tests/chat/goldens/Qwen-Qwen2.5-7B-Instruct-tool_use.txt
@@ -21,7 +21,7 @@ For each function call, return a json object with function name and arguments wi
 Print a hello world message with python.<|im_end|>
 <|im_start|>assistant
 <tool_call>
-{"name": "ipython", "arguments": "{\"code\": \"print('Hello, World!')\"}"}
+{"name": "ipython", "arguments": {"code": "print('Hello, World!')"}}
 </tool_call><|im_end|>
 <|im_start|>user
 <tool_response>
@@ -33,7 +33,7 @@ Anything else?<|im_end|>
 Test a tautology.<|im_end|>
 <|im_start|>assistant
 <tool_call>
-{"name": "test", "arguments": "{\"condition\":true}"}
+{"name": "test", "arguments": {"condition": true}}
 </tool_call><|im_end|>
 <|im_start|>user
 <tool_response>
@@ -45,7 +45,7 @@ Truth is definitely true.<|im_end|>
 Check it on the web.<|im_end|>
 <|im_start|>assistant
 <tool_call>
-{"name": "brave_search", "arguments": "{\"query\": \"what is truth anyway am I right?\"}"}
+{"name": "brave_search", "arguments": {"query": "what is truth anyway am I right?"}}
 </tool_call><|im_end|>
 <|im_start|>user
 <tool_response>

diff --git a/tests/chat/goldens/Qwen-Qwen2.5-Math-7B-Instruct-tool_use.txt b/tests/chat/goldens/Qwen-Qwen2.5-Math-7B-Instruct-tool_use.txt
@@ -21,7 +21,7 @@ For each function call, return a json object with function name and arguments wi
 Print a hello world message with python.<|im_end|>
 <|im_start|>assistant
 <tool_call>
-{"name": "ipython", "arguments": "{\"code\": \"print('Hello, World!')\"}"}
+{"name": "ipython", "arguments": {"code": "print('Hello, World!')"}}
 </tool_call><|im_end|>
 <|im_start|>user
 <tool_response>
@@ -33,7 +33,7 @@ Anything else?<|im_end|>
 Test a tautology.<|im_end|>
 <|im_start|>assistant
 <tool_call>
-{"name": "test", "arguments": "{\"condition\":true}"}
+{"name": "test", "arguments": {"condition": true}}
 </tool_call><|im_end|>
 <|im_start|>user
 <tool_response>
@@ -45,7 +45,7 @@ Truth is definitely true.<|im_end|>
 Check it on the web.<|im_end|>
 <|im_start|>assistant
 <tool_call>
-{"name": "brave_search", "arguments": "{\"query\": \"what is truth anyway am I right?\"}"}
+{"name": "brave_search", "arguments": {"query": "what is truth anyway am I right?"}}
 </tool_call><|im_end|>
 <|im_start|>user
 <tool_response>

diff --git a/tests/chat/goldens/meta-llama-Llama-3.2-3B-Instruct-tool_use.txt b/tests/chat/goldens/meta-llama-Llama-3.2-3B-Instruct-tool_use.txt
@@ -92,23 +92,23 @@ Respond in the format {"name": function name, "parameters": dictionary of argume
 
 Print a hello world message with python.<|eot_id|><|start_header_id|>assistant<|end_header_id|>
 
-{"name": "ipython", "parameters": "{\"code\": \"print('Hello, World!')\"}"}<|eot_id|><|start_header_id|>ipython<|end_header_id|>
+{"name": "ipython", "parameters": {"code": "print('Hello, World!')"}}<|eot_id|><|start_header_id|>ipython<|end_header_id|>
 
 "{\"stdout\": \"Hello, World!\"}"<|eot_id|><|start_header_id|>assistant<|end_header_id|>
 
 Anything else?<|eot_id|><|start_header_id|>user<|end_header_id|>
 
 Test a tautology.<|eot_id|><|start_header_id|>assistant<|end_header_id|>
 
-{"name": "test", "parameters": "{\"condition\":true}"}<|eot_id|><|start_header_id|>ipython<|end_header_id|>
+{"name": "test", "parameters": {"condition": true}}<|eot_id|><|start_header_id|>ipython<|end_header_id|>
 
 "true"<|eot_id|><|start_header_id|>assistant<|end_header_id|>
 
 Truth is definitely true.<|eot_id|><|start_header_id|>user<|end_header_id|>
 
 Check it on the web.<|eot_id|><|start_header_id|>assistant<|end_header_id|>
 
-{"name": "brave_search", "parameters": "{\"query\": \"what is truth anyway am I right?\"}"}<|eot_id|><|start_header_id|>ipython<|end_header_id|>
+{"name": "brave_search", "parameters": {"query": "what is truth anyway am I right?"}}<|eot_id|><|start_header_id|>ipython<|end_header_id|>
 
 "{\"title\":\"Truth: don't ask the web, ask an LLM instead!\",\"url\":\"https://en.wikipedia.org/wiki/Truth\"}"<|eot_id|><|start_header_id|>assistant<|end_header_id|>
 

diff --git a/tests/update_jinja_goldens.py b/tests/update_jinja_goldens.py
@@ -138,7 +138,7 @@ def handle_chat_template(model_id, variant, template_src):
         render_context = json.loads(json.dumps(context))
 
         # Work around Llama-3.1 template quirk: it expects tool_call.function.arguments to be an object rather than its JSON string representation.
-        if 'tool_call.arguments | items' in template_src:
+        if 'tool_call.arguments | items' in template_src or 'tool_call.arguments | tojson' in template_src:
             for message in render_context['messages']:
                 if 'tool_calls' in message:
                     for tool_call in message['tool_calls']: