feat(record): Int6 STE + NorMuon + SWA + Sliding Window (val_bpb=1.16019)

[dexhunter]

Summary

New SOTA submission: mean val_bpb = 1.16019 across 3 seeds, beating current merged SOTA (1.17475) by 0.01456.

Seed	val_bpb	Steps	ms/step	Artifact
1337	1.16146	12357	48.55	15,045,740
42	1.15935	12351	48.58	15,053,489
7	1.15976	12336	48.69	15,157,415
Mean	1.16019

Key Techniques

1. Int6 STE — Fake int6 per-row quantization ([-31,31]) on every forward pass with straight-through gradient bypass. Model learns to cope with quantization noise, yielding only ~+0.002 bpb gap.
2. NorMuon optimizer — Row-normalized Newton-Schulz updates on top of Muon with adaptive step sizes.
3. 3x MLP width (1536) — Enabled by int6 compression savings within 16MB budget.
4. FP16 tied embedding — Embedding tensor stored in fp16, never quantized.
5. Sliding window eval (stride=64) — Every token gets 960 tokens context (~0.033 bpb improvement).
6. SWA — 7 checkpoint average during warmdown phase.
7. Zstd-22 compression — Better ratio than zlib for quantized weights.
8. U-Net skip connections — Encoder-decoder structure with learnable skip weights.

Architecture

• 9 layers, 512 dim, 8 heads, 4 KV heads (GQA)
• Vocab 1024 (SentencePiece BPE), seq len 1024
• relu² activation, RoPE, logit softcapping (30.0)

Submission checklist

• 3-seed verification with mean val_bpb
• All artifacts < 16MB
• Wallclock < 600s on 8xH100
• Train logs included
• Reproducible train_gpt.py included
• submission.json with metadata

[MatoTeziTanka]

Community Review — feat(record): Int6 STE + NorMuon + SWA + Sliding Window (val_bpb=1.16019)

BPB: 1.16019 | Compliance: LOOKS CLEAN — pure-neural submission, no TTT/SLOT/n-gram-cache

What I found in the code (head SHA 2c12ff54420b, file records/track_10min_16mb/2026-03-20_Int6STE_NorMuon_SWA_SlidingWindow/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 0.03s, dim=512, layers=9, vocab=1024, code=54630 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 0.03s, dim=512, layers=9, vocab=1024, code=54630 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.