Qwen3-4B-V2-Superior-Hybrid-30k

What this model is

• Base model: Qwen3-4B (original hybrid-thinking)
• Dataset: superior_hybrid 30k
• Training method: Single-stage SFT (V2 architecture)

Notes

V1 vs V2 (architecture explanation)

• V1: Registers /think and /no_think as new single special tokens. Qwen3-Base does not recognize these tokens, so their embeddings are randomly initialized. This shadows Qwen3's original think/no_think control ability — both experts start from the "think" mode.
• V2: Uses the original Qwen3 tokenizer without registering new tokens. /think and /no_think are encoded as multi-token sequences via BPE. The router detection scans the trailing window of the input for characteristic tokens (5854 = no_think, 26865 / 20439 = think). This preserves Qwen3's native understanding capability — the two experts start from Qwen3's already-trained think/no_think modes.