exp190 — saturating soft data without starving training
Problem
exp186 soft cache was ~56% opening / 22% mid / 22% end. Openings label faster (shallower trees) so naive HF streaming floods the dataset. Training then overfits early-game priors and underperforms in middlegame/endgame — exactly where Elo is decided without MCTS.
Data architecture
| Lever | Design |
|---|---|
| Phase quotas | 22% opening / 48% middlegame / 30% endgame (hard gate at writer) |
| Depth-by-phase | op 10–14, mid 12–16, eg 14–18 (full-strength SF18, no Elo limit) |
| Multi-source FENs | HF stream + book playouts + endgame templates + random walks |
| Deficit-first feed | Producer oversamples the lagging phase every cycle |
| Syzygy | Wired into each worker when syzygy/*.rtbw present |
| soft_cache tags | phase + label_depth tensors for stratified training |
Training efficiency (how to consume this)
- Stratified batches — sample 1/3 from each phase bucket every step so gradients never bog on openings.
- Depth weights —
w = label_depth / mean_depth(or clip 0.5–1.5) so deeper MultiPV pulls harder than shallow noise. - Critical upweight — boost high
cp_gap_top1_top2/ low entropy rows (forcing moves teach more than quiet equals). - Merge, don't replace — mix exp190 deep cache with exp186 2M shallow at ~30/70 so you keep volume + add quality.
- Round-robin curriculum — alternate soft-deep / soft-shallow / hard-HF micro-batches inside the same step (already partially done via soft_frac).
Model-side advantages (no MCTS)
- Phase-balanced policy → fewer “opening genius / endgame idiot” failures
- Deeper soft targets → better next-move calibration under legal-mask argmax
- Endgame + Syzygy teacher → converts tablebase-perfect moves into policy mass
- Compact vocab + spatial head already amortizes move representation
Ops
- SF binary:
stockfish/stockfish-latest→ SF18x86-64-vnni512(Xeon 6342) - 64 workers × 96MB hash ≈ 6GB; ~441GB RAM free on this box
- Safe alongside GPU exp189 (CPU-only)
- Tail:
tail -f outputs/exp190_phase_deep/run.log