e2rea1/RouteWeaver-4B
Text Generation • 4B • Updated • 81
RouteWeaver
One 4B router that picks an execution mode, then routes six frozen workers inside it. Code + paper: github.com/LaughKing/RouteWeaver
Note The reported router. Qwen3-4B-Instruct-2507, all parameters fine-tuned, task reward only (alpha = 0).
e2rea1/RouteWeaver-data
Viewer • Updated • 2.85k • 15Note The exact training and evaluation sets as verl prompt rows, with the .meta.json sidecars training needs.
e2rea1/RouteWeaver-4B-cost
Text Generation • UpdatedNote Cost-aware variants: alpha = 0.1 / 0.3 / 0.5, one subdirectory each. Same recipe, efficiency reward mixed in.