Fork of geodesic-research/nemotron-think-tokenizer @ cce18a60d4e17b9a0436706fbdc17b706994270b with an identical encoder and three chat-template edits:

Net effect: rendering never strips prior-turn reasoning (truncate_history_thinking defaults to False) and never emits an empty <think></think> stub for assistant messages without reasoning. The {% generation %} loss-mask markers are unchanged.

Built 2026-08-09 21:57 UTC by scripts/data/build_think_history_tokenizer.py (geodesic-megatron).