A model soup (0.5/0.5 arithmetic weight average, no training) of two dense Qwen3-4B SFT checkpoints trained on the same 30k data, differing only in thinking mode:

ShourenWSR/Qwen3-4B-Dense-NoThink-30k (no_think-only SFT)
ShourenWSR/Qwen3-4B-Dense-Think-30k (think-only SFT)

Purpose

Single-model naive baseline ("merged-soup") in the PL-MoE rebuttal 4-way comparison: no_think-only / think-only / merged-soup / PL-MoE.

Merge details

Per-tensor arithmetic mean, weights 0.5 / 0.5
Accumulation in float32, cast back to original bf16
Config + tokenizer copied verbatim from the no_think parent (both parents byte-identical)
399 tensors averaged; key sets / shapes / dtypes verified identical before merge

ShourenWSR/Qwen3-4B-Dense-Merged-Soup-30k

Qwen3-4B-Dense-Merged-Soup-30k

Purpose

Merge details

ShourenWSR/
Qwen3-4B-Dense-Merged-Soup-30k