LAWA-EMA frontier fork (pr198 base, SWA -> LAWA val_bpb=1.1551)

[machdragon]

Summary

• val_bpb = 1.1551 (int6 sliding window, stride=64) | 12.7 MB artifact | 8xH100
  SXM, 600s
• Based on PR 11-Layer Int6 + WD=0.04 + SWA + FA3 (val_bpb: 1.1318) #198 (11L, int6, MLP3x, relu², FA3, SmearGate, BigramHash, OrthoInit,
  U-Net skips, WD=0.04)
• SWA replaced by LAWA-EMA (exponential moving average, decay=0.995, float32
  shadow, every-step update)
• Overtone init added (SVD power-law embedding spectrum for smoother int6
  quantization)
• Two bug fixes: bigram proj zero-init override, sliding window partial-window overlap

Metric	PR #198	This
Int6 sliding val_bpb (s64)	1.1318	1.1551
Int6 roundtrip val_bpb	1.1543	1.1779
Artifact size	15.7 MB	12.7 MB
Steps (600s)	7,412	6,715

Single seed run (seed=1337). Additional seed runs pending for statistical validation.

Test plan

• Maintainer re-evaluation of train_gpt.py on 8xH100 with official harness
• Multi-seed validation (seeds 1337, 42, 2025)

Our numbers:

┌────────────────────────────┬────────┐
│ Metric │ Value │
├────────────────────────────┼────────┤
│ Pre-quant val_bpb │ 1.1622 │
├────────────────────────────┼────────┤
│ Int6 roundtrip val_bpb │ 1.1779 │
├────────────────────────────┼────────┤
│ Int6 sliding val_bpb (s64) │ 1.1551 │
└────────────────────────────┴────────┘

Note: our 1.1551 is worse than PR #198's 1.1318. The artifact is 3 MB smaller (12.7 vs
15.7 MB), but the BPB regressed.

[MatoTeziTanka]

Community Review — LAWA-EMA frontier fork (pr198 base, SWA -> LAWA val_bpb=1.1551)

BPB: 1.1551 | Compliance: LOOKS CLEAN — pure-neural submission, no TTT/SLOT/n-gram-cache

What I found in the code (head SHA 31c16a6abb90, file records/track_10min_16mb/2026-03-20_LAWA_EMA_Int6_MLP3x_OvertoneInit/train_gpt.py):

Static code review found no TTT adaptation function, no SLOT optimization loop, no n-gram-cache class, and no pre-quant val-token fine-tune. The eval path uses the standard sliding-window stride-64 pattern. The submission is a pure-neural architecture iteration on the standard SP1024/SP4096/SP8192 baseline.

CPU smoke test (CT2038 proteus-engine, 2026-04-11): import OK in 0.08s, dim=512, layers=9, vocab=1024, code=65258 B, SMOKE_TEST_PASS

Verdict: LOOKS CLEAN.

Recommendation to @cocohearts @valerio-oai @0hq @yuzhougu-oai @notapplica: MERGE pending the usual record-track checks (3-seed validation, under-16MB artifact cap, ≤600s train + ≤600s eval on 8×H100 SXM). No compliance flags from the classification pass — this looks like a clean pure-neural iteration on the standard baseline.

Auto-classification caveat: this review was drafted by the AST-based classifier. If there's a non-standard eval mechanism (logit postprocessing, hedge mixing, etc.) that I missed because it's factored into a helper file or a non-standard function name, please flag it and I'll re-run the audit manually.

---

Reviewed by @MatoTeziTanka — The Agora. CPU smoke test (CT2038 proteus-engine, 2026-04-11): import OK in 0.08s, dim=512, layers=9, vocab=1024, code=65258 B, SMOKE_TEST_PASS. Classification via deterministic AST-based classify_prs.py (pattern bank derived from ~65 manually-reviewed PRs earlier in the 2026-04-11 sweep). This review was auto-drafted from a template and spot-checked before posting — if the template misread your code, please call it out so I can iterate the classifier.