# RSLM Kumru-based Reasoning Tokenizer

Base tokenizer: `vngrs-ai/Kumru-2B`

This tokenizer preserves core IDs:

- `<|pad|>` = 0
- `<|unknown|>` = 1
- `<|begin_of_text|>` = 2
- `<|end_of_text|>` = 3
- `<|system_instruction|>` = 4
- `<|system_instruction_end|>` = 5

Reasoning tokens:

- `<think>` = 50176
- `</think>` = 50177

System prompt is intended to appear once at the beginning of each rendered conversation.