es-spirals — evolution-strategies training from scratch

The gradient-free counterpart of ../spirals, and the from-random-init acceptance test of fucina.es: the same two-hidden-layer tanh MLP (hidden 64, 194 points), same data generator, but no backward pass and no optimizer — reward is -CE on the full batch with z-score shaping and mirrored (antithetic) sampling on the ES-at-scale update (main.zig).

It is also the member-parallel showcase: evaluateMembers fans the population over worker threads, each owning a full MLP replica and its own ExecContext; materializeMember writes theta + sigma*eps into the replica and only scalar rewards come back. The defaults reach 100% accuracy within ~15k iterations (~75 s ReleaseFast on an M1 Max, per the source header).

Self-verifying: exits nonzero unless the trained network reaches --target accuracy (default 0.90) on the training set — chance is 0.50.

The same trainer fine-tunes a real LLM gradient-free: ../es_finetune.

zig build es-spirals -Doptimize=ReleaseFast
# all knobs:
zig build es-spirals -Doptimize=ReleaseFast -- [--iterations N] [--population N] \
    [--sigma F] [--alpha F] [--workers N] [--norm z_score|centered_ranks] [--seed N] [--target F]

Defaults: --iterations 15000, --population 128 (antithetic when even), --sigma 0.1, --alpha 0.1, --workers 4, --norm z_score, --seed 42, --target 0.9. z-score is the deliberate default: with a bounded well-behaved reward its magnitude information converges where centered_ranks stalls — pick the shaping by reward regime (docs/TRAINING.md §13).

Quick sanity check

The run is deterministic per --seed, so a shorter run is a prefix of the default one; the default seed sits at 100% train accuracy from ~iteration 6500, so about half the default iteration budget already clears the 0.90 gate:

zig build es-spirals -Doptimize=ReleaseFast -- --iterations 7000

Expect a before: accuracy near chance, an iter … accuracy … ce … progress line every 500 iterations, and a final PASS: gradient-free from-scratch training reached 100.0% (target 90.0%) with exit code 0.

Shared knobs

The ReleaseFast/-Dcpu build discipline and global thread/BLAS knobs are shared machinery — see docs/RUNNING-MODELS.md. This demo needs no model weights.