6L Depth Minimalism + U-Net + Sliding Window — val_bpb 1.2026

[alphastar1111]

Summary

• 6-layer transformer (3 encoder + 3 decoder) matches the 9L baseline with 33% fewer layers and fewer
  parameters
• Sliding window val_bpb: 1.2026 (post-quant standard: 1.2246 vs baseline 1.2244)
• Artifact size: 15.84 MB (under 16MB)
• Estimated 8xH100 runtime: ~8.2 min (trained/validated on 1xA100, calibrated via baseline step_avg ratio
  705.97ms/43.54ms = 16.2x)

Approach

Five architectural bets compensate for extreme depth reduction:

1. Full attention (4H/4KV, no GQA)
2. Untied embeddings
3. Half batch (262K tokens, grad_accum=2) — 2x more optimizer steps per token
4. Tight logit softcap (12.0 vs 30.0) — implicit regularization for shallow models
5. Long context (seq_len=2048) — double the baseline's 1024

Credits

• LeakyReLU^2 (slope=0.5) — from PR Record: LeakyReLU² + Legal Score-First TTT + Parallel Muon — val_bpb 1.1194 (3-seed mean) #549 by @abaybektursun
• Sliding window eval (stride=64) — from @mattqlf's
  records/track_10min_16mb/2026-03-19_SlidingWindowEval/README.md submission
• seq_len=2048 — first explored by @spokane Way in
  records/track_10min_16mb/2026-03-18_LongContextSeq2048/README.md
• U-Net skip connections — widely adopted across leaderboard (PR 11-Layer Int6 + WD=0.04 + SWA + FA3 (val_bpb: 1.1318) #198 and derivatives)
• MLP 3x expansion — standard from PR 11-Layer Int6 + WD=0.04 + SWA + FA3 (val_bpb: 1.1318) #198 / @raahil Shah's Int6 MLP3x submission
• Muon optimizer — from the modded-nanogpt baseline

Hardware details

Trained on 1xA100 80GB (not 8xH100). Calibration math and caveats about small-batch GPU inefficiency are
detailed in the README. MAX_WALLCLOCK_SECONDS=600 warmdown handles graceful cutoff if timing estimate is
off.

Test plan

• Full 20K step training run on 1xA100
• Int8+zlib roundtrip validation passes
• Sliding window eval (stride=64) completes
• Artifact under 16MB (15,841,206 bytes)
• Code under 1500 lines (1,325 lines)
• Verify on 8xH100 (extrapolated from A100 calibration)

[MatoTeziTanka]

Community Review — 6L Depth Minimalism + U-Net + Sliding Window — val_bpb 1.2026

BPB: 1.2026 | Compliance: LOOKS CLEAN — pure-neural submission, no TTT/SLOT/n-gram-cache

What I found in the code (head SHA 27c66e4125ee, file records/track_10min_16mb/2026-04-10_6L_DepthMinimalism_UNet_SlidingWindow/train_gpt_26e6b4a.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=6, vocab=1024, code=56975 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=6, vocab=1024, code=56975 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.