dualbrain-bihmoe-poc

Minimal PoC for bilateral hierarchical MoE with reconciliation. Goal: determine if structure yields OOD/generalization signal vs dense compute-matched baseline.

Protocol:

Deterministic task generation (seeded).
Side-by-side training (structured S vs dense D_a) from the same data stream.
Eval every N steps on fixed IID/OOD/structure-break sets.
Early-kill criteria documented in docs/00_north_star.md.

Reproduce (twohop_bind, Schedule-2+Ramp+BraidMix @ 4k)

Environment (Arch, system torch-cuda; venv uses system site-packages):

uv venv --python /usr/bin/python --system-site-packages
uv sync

Run 3-seed panel with live KEYLINE progress:

scripts/run_panel_4k.sh configs/poc_twohop_sched2_ramp_braid.yaml

Extract KEYLINEs from logs:

scripts/extract_keylines.sh /tmp/bihmoe_s11_4k_*.log

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support