FINAL-Bench/Darwin-Chimera-4B-MoE-2E-PoC

🤗 Hugging Face 来源text-generationapache-2.08.7B 参数17 GBsafetensors✓ 2 个校验和今天更新

一条命令提交

在你的模型文件夹旁边运行它。它会制作种子、将文件与 Hugging Face 比对，然后提交。你只需开始做种，并粘贴你账户中的密钥。它只读取你的文件，绝不修改。如果愿意，可以先阅读脚本。

curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo FINAL-Bench/Darwin-Chimera-4B-MoE-2E-PoC ./model-folder

Darwin-Chimera-4B-MoE-2E-PoC (private)

2-expert MoE upcycling PoC: Qwen3.5-4B (expert0) + Reasoning-Fix (expert1), fused as one Qwen3.5-MoE with a trained router (frozen experts).

Two parents' FFNs preserved 100% as frozen experts (trace intact, single-parent non-reducible).
Router trained (163K params) → both experts genuinely used (~49-56%), coherent generation.
Honest result: neutral vs base on accuracy (parents too similar) — this is a mechanism/originality PoC, not a performance model. Real gains come from divergent parents + RFT-edge (see Darwin-V9-Chimera-4B-RFT).

Build/verify scripts: chimera_build.py / chimera_router_train.py / chimera_verify.py.