Record: Int6 STE + SmearGate + Seq2048 + OrthoInit + RoPE50K + SWA/100 (mean val_bpb=1.1507)

[dexhunter]

Summary

Mean val_bpb = 1.1507 (3-seed verified, p<0.001), beating merged SOTA (1.1748) by 0.024.

Evolved over 31 AIDE2 optimization steps from baseline 1.1607 on 8xH100.

Seed	val_bpb	Steps	ms/step	Artifact
1337	1.15022	10613	56.53	14,555,057
42	1.15095	10610	56.53	14,791,593
7	1.15099	10610	56.53	14,562,412
Mean	1.15072

Technique Stack

1. Int6 STE — Fake int6 quantization every forward pass with STE gradient bypass
2. NorMuon + WD=0.02 — Row-normalized Newton-Schulz with decoupled weight decay
3. 3x MLP (1536 hidden) — Wider MLP enabled by int6 compression
4. SmearGate — Learned gate blending token embeddings with predecessors (~512 params)
5. Orthogonal Init — OrthoInit on all non-zero-init linear layers
6. Seq2048 + RoPE Base 50K — 2x training context with adjusted RoPE
7. SWA every 100 steps — More frequent checkpoint averaging during warmdown
8. FP16 tied embedding — Embedding never quantized
9. Sliding window eval (stride=64) — Every token scored with ~1984 tokens context
10. Zstd-22 — Better compression than zlib
11. U-Net skip connections — Encoder-decoder with learnable skip weights

Architecture

• 9 layers, 512 dim, 8 heads, 4 KV heads (GQA)
• Vocab 1024 (SentencePiece BPE), seq len 2048, tied embeddings
• relu² activation, RoPE, logit softcapping (30.0)

Submission checklist

• 3-seed verification (mean val_bpb=1.1507)
• All artifacts < 16MB (max 14.79MB, 1.2MB headroom)
• Wallclock < 600s on 8xH100
• Train logs included (3 seeds)
• Reproducible train_gpt.py included
• submission.json with metadata

[MatoTeziTanka]

Community Review — Record: Int6 STE + SmearGate + Seq2048 + OrthoInit + RoPE50K + SWA/100 (mean val_bpb=1.1507)

BPB: 1.1507 | Compliance: LOOKS CLEAN — pure-neural submission, no TTT/SLOT/n-gram-cache

What I found in the code (head SHA fb1b0345b4dd, file records/track_10min_16mb/2026-03-20_Int6STE_SmearGate_Seq2048_OrthoInit_RoPE50K_SWA100/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 0.23s, dim=512, layers=9, vocab=1024, code=55785 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 0.23s, dim=512, layers=9, vocab=1024, code=55785 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.