Record: SmearGate BOS Fix 3-Seed Reproduction — val_bpb 1.06145 (3-seed mean)

[Christopher-Lee-McClendon]

Record: SmearGate BOS Fix — 3-Seed Reproduction of PR #1851

val_bpb = 1.06145 (3-seed mean, std 0.00068) | ~15.95 MB | 8×H100 SXM 80GB

Summary

Pure reproduction study of PR #1851 by @aquariouseworkman. The training script is byte-identical to the code in #1851. No new techniques or modifications are introduced.

PR #1851 submitted a single-seed result (seed 42, val_bpb = 1.06128). This PR extends that to a 3-seed evaluation (seeds 42, 314, 1234) confirming the result is robust and reproducible.

3-Seed Results

Seed	Pre-Quant BPB	Quant BPB	Post-TTT BPB	Artifact (bytes)	Train Time	Eval Time
42*	1.06490240	1.07405660	1.06128183	15,952,086	599.6s	519.5s
314	1.06467893	1.07358634	1.06086831	15,952,419	599.6s	525.6s
1234	1.06593114	1.07503808	1.06220261	15,952,690	599.5s	479.6s
Mean ± Std			1.06145 ± 0.00068

* Seed 42 result from original PR #1851 author @aquariouseworkman. Seeds 314 and 1234 are independent runs by @Christopher-Lee-McClendon.

Technique Stack

All techniques inherited from #1851 (and its lineage):

Technique	Source	Author
Base architecture (11L, MLP 4x, MuonEq-R)	PR #1787	@nprime06
SmearGate attention	PR #1797	@dexhunter
SmearGate BOS fix	PR #1851	@aquariouseworkman
LQER Asymmetric quantization	PR #1797	@dexhunter
CaseOps SP8192	PR #1729	@romeerp
GPTQ + SP8192	PR #1394	@clarkkev
Score-first TTT (3 phases)	PR #549	@abaybektursun
BOS bug identification	Issue	@cocohearts

Compliance

Budget	Limit	Worst-Case	Status
Artifact size	16,000,000 bytes	15,952,690 bytes	✅
Training time	600s	599.6s	✅
Eval time	600s	525.6s	✅

Reproduction

# Install deps
pip install brotli python-minifier

# Prepare CaseOps SP8192 data
python3 prepare_caseops_data.py  # from romeerp/parameter-golf-caseops-v1

# Train (replace SEED)
SEED=42 CASEOPS_ENABLED=1 EMBED_BITS=7 SMEAR_GATE_ENABLED=1 \
SPARSE_ATTN_GATE_ENABLED=1 MIN_LR=0.1 EMBED_CLIP_SIGMAS=15.0 \
MLP_CLIP_SIGMAS=12.0 GPTQ_RESERVE_SECONDS=0.5 PHASED_TTT_NUM_PHASES=3 \
torchrun --standalone --nproc_per_node=8 train_gpt.py

Hardware: 8×H100 SXM 80GB (RunPod, matotezitanka/proteus-pytorch:community image)

Credits

• @aquariouseworkman — PR Record: val_bpb = 1.06128 SmearGate BOS Fix + PR #1787 Base + Smear Gate + LQER Asymmetric + Phased TTT (indirect 3 seed mean) #1851 author (SmearGate BOS fix, seed 42 result)
• @nprime06 — PR Record: PR #1736 + Polar Express NS + MIN_LR + Sparse Attn Gate + Fused CE + PR #1767 TTT — val_bpb 1.06335 #1787 (base architecture)
• @romeerp — PR Record: CaseOps Tokenizer + Tapered WD - val_bpb 1.0678 (3-seed mean) #1729 (CaseOps)
• @dexhunter — PR Record: PR #1787 base + Smear Gate + LQER Asym — val_bpb 1.06157 #1797 (SmearGate + LQER)
• @cocohearts — BOS boundary bug identification
• @abaybektursun — PR Record: LeakyReLU² + Legal Score-First TTT + Parallel Muon — val_bpb 1.1194 (3-seed mean) #549 (score-first TTT)
• @clarkkev — PR Record: SP8192 + GPTQ Embeddings + Depth Recurrence + MuonEq-R + SDClip — val_bpb 1.08563 (5 seed mean) #1394 (GPTQ + SP8192)

[aquariouseworkman]

You are amazing!!!