Hi, first of all — I really love this model. The humanlike tone is noticeably more natural than the base Qwen for companion / casual chat scenarios, and it genuinely feels like talking to a person rather than an assistant. Great work on the fine-tune! Quick question though — is this model not suitable for tool/function calling? I'm using it via llama.cpp in an agent framework, and it often answers fact-based questions without calling the attached tools (even though a direct curl shows the server does return tool_calls correctly). Wondering if this is expected from the humanlike fine-tune, or if there's a recommended setup to improve tool-use reliability. Thanks!
Thanks for the kind words! The GGUF has a tool-call template, and your curl test shows llama.cpp can return tool calls. But this fine-tune was optimized for natural conversation, not tested for reliable tool selection, so it may answer directly when tool_choice is auto.
Use --jinja --reasoning off, then compare the exact request your framework sends with the working curl request. If a lookup is essential, require it in your app or try tool_choice: "required" if your setup supports it. If you share a redacted request, llama.cpp version, and quant, I can help narrow it down.
