On-device intent router for the TalkPlay hands-free YouTube voice player.
Maps a spoken Korean/English utterance directly to a single TalkPlay intent
JSON — no system prompt required; the tool definitions live in the
fine-tuned weights.
Base: Gemma 3 1B instruct (bf16), LoRA fine-tuned on 1,638 synthesized
utterances covering all 78 TalkPlay intents (ko/en, STT-style variants),
then fused and post-training quantized to 8-bit (group size 64).
Eval: 54/54 (100%) on the shared TalkPlay router suite; ~0.4 s per
command on M-series hardware (vs ~6.8 s for the previous Gemma 3 4B +
1,400-token-prompt setup).
Usage: send the raw utterance as the user message with the bundled chat
template. Output is one JSON object, e.g.
{"intent": "SEARCH", "query": "calm jazz"}.
Training pipeline: Tools/agent-router in the TalkPlay repository.
