avewright/chess-moe-router

Frozen-expert MoE router

Small MLP gate over five frozen 99M squares64 specialists. Only this router is trained; the experts stay in their own repos.

This file is latest.pt at router step 550 (2026-09-14 19:08 UTC). Train loss ~0.6869. Val source-cls CE ~0.4296. Val acc ~0.839.

Not a 99M policy. Not the incumbent, puzzle, endgame, opening, or middlegame expert.

Experts

• incumbent → avewright/chess-transformer-100m-squares64
• puzzle → avewright/puzzle-model
• endgame → avewright/endgame-model
• opening → avewright/opening-model
• middlegame → avewright/middlegame-model

Architecture

• Stem: frozen incumbent global_hidden (736d)
• Head: LayerNorm → Linear(736,256) → GELU → Linear(256,256) → GELU → Linear(256,5)
• Params: 257,221
• Vocab: compact 1968
• Dispatch: hard argmax, one specialist forward (incumbent encode is reused if chosen)

Files

• latest.pt — router weights + expert repo list
• step_000550.pt
• router_config.json
• train.log