LGAI-EXAONE

/

K-EXAONE-236B-A23B

like 557

Follow

LG AI Research 1.66k

Hello, @lesj0610 . Thank you for your attention.

We checked your case but could not reproduce the same failure.

Here is the snippet we tested:

tools = [{
    "name": "get_current_weather",
    "description": "Get the current weather for a given city",
    "parameters": {
        "type": "object",
            "properties": {
                "location": { "type": "string", "description": "City name with country code (e.g., Seoul,KR)" },
                "unit": { "type": "string", "enum": ["metric", "imperial"] }
            },
        "required": ["location"]
    }
}]

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "지금 서울의 현재 기온이 몇 도인가요?"},
]

inputs = tokenizer.apply_chat_template(
    messages,
    return_tensors="pt",
    add_generation_prompt=True,
    tools=tools,
    chat_template_kwargs={"enable_thinking": False},
)

output = model.generate(
    **inputs,
    do_sample=True,
    temperature=1.0,
    top_p=0.95,
    max_new_tokens=256,
)

Please note that the tools schema should be passed to the HF tokenizer, not directly to model.generate(), when using a tool-augmented chat template.
(For inference server, please check the documentation of using OpenAI-compatible tool calling. )

With this setup, we get the following tool call:

<tool_call>{"name": "get_current_weather", "arguments": {"location": "Seoul,KR", "unit": "metric"}}</tool_call>

It seems the tool schema might not be passed to model in proper way.
If you can share the generation part of your code, we can take a closer look.

Thanks for the clarification and the working snippet.

One key detail: my testing was done only via the FriendliAI web UI (I cannot run the 236B model locally). In that UI, I do not have any way to verify whether the tools schema is actually injected into the model inputs (i.e., whether the server uses a tool-augmented chat template like tokenizer.apply_chat_template(..., tools=tools, ...)).

So the failure I observed (API-style JSON explanation instead of a tool call) may be due to the FriendliAI web UI pipeline not passing tools/schema, rather than the base model lacking tool-call training.

Could you confirm whether the FriendliAI web UI currently supports tool calling with schema injection?
If yes, could you share the minimal steps/settings in the UI to enable it, or provide an OpenAI-compatible endpoint example request that triggers tool calls?

If the web UI does not inject tools/schema, then my observation is expected from the deployment, and I will update my report accordingly.

Thank you.

Thanks for the clarification. I was testing via the FriendliAI web UI, where I couldn’t verify whether the tools schema was actually injected. After switching to the FriendliAI OpenAI-compatible API and explicitly passing tools (and tool_choice), the model correctly returned finish_reason: "tool_calls" with a valid get_current_weather tool call. I’ll update my report accordingly.