Federated training of a 2-layer MLP head (D=384 → H=128 → K=4, P = 49 796)
on pre-computed MiniLM-L6-v2
features for 4-class AG News topic classification (World / Sports /
Business / Sci-Tech). The features are baked into a 154 KB static
asset (/data/agnews-mini.bin) — no MiniLM inference in the
browser. Each tab is a federated SPSA worker; the server uses
Phase 39's symmetric AIMD adaptive η.
Protocol: SPSA tournament (DeComFL fusion, Phase 36+).
Each round, this worker runs 4 SPSA trials on its private slice of
25 training examples; submits the trial with the most-negative
claimed Δloss as a 20-byte proposal (seed, scalar_g, delta).
Server picks the best across all workers, applies
θ ← θ − η · g · u(seed), broadcasts the applied flip.
Adaptive η: server uses Phase 39's symmetric AIMD —
η ← η × 1.05 on a real loss drop, ← η × 1/1.05 on a regression,
computed from the server's own test-set forward pass on the 25
held-out examples.
Federation: 3 disjoint training shards (25 examples each).
Your shard is fixed by the hash of your worker id. With 3+ joined
peers the swarm covers all training data; the server scores on the
shared test set (the byzantine truth).