Luigi/nemotron-3.5-asr-streaming-0.6b-zhtw

v3: teacher-filtered data + two-phase repair — CV zh-TW 12.03->9.81 (18% better); small fr/hi/en cost, ar/ja/ko/de still beat base

model card: prompt-slot guidance — the FT repairs `auto` (base 50.58 -> 11.33 CER, best slot); all slots within ~0.7 CER

model card: full 9-eval table — ja/ko/hi/de/en all improve, only fr/es regress ~1.5

v2: zh-TW fine-tune (CV zh-TW CER 38.43->12.03, 3.2x better); en/de improve, fr/es +1.5/+1.3, ar flat. Trained on Simplified targets (tokenizer cannot represent Traditional) — use OpenCC s2t.

zh-TW warm-started variant (v1.1): lossless zh-CN->zh-TW prompt column copy; transformers-loadable

initial commit