myclew

Mirror of the fleet agent codebase — reusable, config-driven Kaggle agents + their data-wise verifiers — plus the per-competition experiment code built on top of them.

• fleet_agents/ — the agents (BaseAgent subclasses): tabular/GBM, geology particle-filter (Track B), neural-sequence Muon trainer (Track C), low-bit QAT (NVFP4/MXFP4/int8), setup-env GPU guard (the verified cu128 / CUDA-12.8 sm_120 stack for the RTX 5090), and many more.
• test_fleet_agents/ — one data-wise verifier per agent (offline, deterministic).
• competitions/<slug>/ — experiment code + YAML configs we build per competition (no data).

Source of truth lives in the biohub competition workspace; this repo is a stable mirror, synced and committed every 3 hours (only when changes are pending, and only if every .py byte-compiles).

Sync manually: ./sync_and_commit.sh