FormosaNLU-Synth: Filtered Synthetic Data Distillation for Traditional Chinese (Taiwan) Natural Language Understanding

This evidence-bounded technical report documents FormosaNLU-Synth, a local-first synthetic-data distillation pipeline for low-resource MASSIVE zh-TW joint intent classification and slot filling. It reports the frozen Gemma experiments, preregistered Phi-4-mini cross-family replication, robustness probes, equal-N recipe ablation, resource accounting, public artifact hashes, and limitations. This archival technical note has not undergone peer review. The corresponding software release is archived at https://doi.org/10.5281/zenodo.21879133.