Submission: 10L + Sliding Window eval (mean val_bpb=1.1899)

[shajalahamedcse]

Key idea:
The model was like a student who studied short paragraphs but was being tested on long chapters — so we asked: what if it practiced on long text too? We changed one line (train_seq_len = 4096) so it trained on 4096-token passages instead of 1024, teaching it real long-range patterns, then evaluated it with overlapping windows (stride=64) so every word gets maximum context during scoring. We ran it on Modal 8×H100 GPUs, got a consistent mean of 1.1899 across 3 random seeds (1337, 42, 7)

Combine train_seq_len=4096 with 10 layers and sliding window evaluation (stride=64)

Seed results

Seed	val_bpb	Artifact size
1337	1.1900	15,115,793 B
42	1.1908	15,128,724 B
7	1.1888	15,154,068 B
mean	1.1899	≤ 16MB ✓
std	0.0008

Config

num_layers     = 10
train_seq_len  = 4096
eval_stride    = 64  (sliding window)
warmdown_iters = 3600
matrix_lr      = 0.04
muon_momentum  = 0.95

Hardware

Modal 8×H100 SXM, torchrun --standalone --nproc_per_node=8, MAX_WALLCLOCK_SECONDS=600

[MatoTeziTanka]

Community Review — Submission: 10L + Sliding Window eval (mean val_bpb=1.1899)

BPB: 1.1899 | Compliance: LOOKS CLEAN — pure-neural submission, no TTT/SLOT/n-gram-cache

What I found in the code (head SHA f8aaba857564, file records/track_10min_16mb/2026-03-20_Seq4096_10L_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 0.07s, dim=512, layers=10, vocab=1024, code=55483 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 0.07s, dim=512, layers=10, vocab=1024, code=55483 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.