gemma-4-E4B-stitch-zh

google/gemma-4-E4B-it fully finetuned (language model; vision tower frozen) on voidful/agent-sft-stitch-zh — a Traditional-Chinese Agent-STITCH-S speech-first voice-assistant dataset.

The model produces speech-first agent trajectories: a safe spoken opening, then interleaved [SOPR]…[EOPR] reasoning, <TOOL_CALL>{…}</TOOL_CALL> tool calls, short spoken "waiting" lines covering tool latency, and a final spoken answer ending with [EOR]. Every spoken chunk is TTS-ready (no code/paths/urls/symbols read aloud; numbers spoken out).

Training

• Full finetune of the language model (7.77B trainable; vision tower frozen), FSDP on 4×H200.
• Loss masked on the prompt and on <TOOL_RESULT> spans (tool results are environment-provided).
• 3 epochs, effective batch 64, lr 1e-5 cosine, max_len 6144. Loss 1.23 → ~0.14.

Inference (STITCH-S loop)

Build the prompt as system (STITCH-S instructions + available tools) + user. Generate until <TOOL_CALL>, execute the real tool, inject <TOOL_RESULT>{…}</TOOL_RESULT>, then continue generating until [EOR].

Data pipeline

Source dialogues were converted from voidful/agent-sft with google/gemma-4-31B-it, scored on an 8-dim/16-pt rubric, and filtered to total_score >= 12 plus a complete tool_call↔result check.