LoRA adapter used by team Hephaestus's LifeMate agent (Vietnamese voice-to-action). Given the system prompt, the JSON schemas of the installed apps' tools, the user context, the conversation history and earlier tool results, the model outputs the next action as one JSON object (a tool call or a final message).
Qwen/Qwen3.5-4B @ 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0atrain split (1,112 samples, 3,517 next-action rows);
no test data, external data or synthetic dataadapter_model.safetensors sha256 a51e4f4e0166f262deaa1d48578701d1481d27016c2a168d0aac8ba7c01a5f58In the submission the adapter is merged into the base model in bf16. Public test score of the full system (with the
companion ASR phuc2k3/lifemate-phowhisper-large-soup3): 63.79.
Training code: training_code/slm_e3 in the team's submission image.
Qwen/Qwen3.5-4B