Record: SP4096 + Linear LR + Depth Recurrence -- val_bpb=1.0924 (3-seed mean)

[dttdrv]

Summary

val_bpb: 1.0924 (3-seed mean, std 0.0004) | 15.99 MB artifact | 8xH100 SXM, 600s

No TTT, no SLOT, no n-gram cache, no eval-time adaptation. All four conditions from Issue #1017 satisfied. Evaluation is pure sliding-window at stride=64.

Improvement over current SOTA (PR #1019, 1.1147 BPB): -0.0223 BPB (Welch t=-68.85, df=3.84, p << 0.001)

Results

Seed	Steps	Pre-quant BPB	Sliding BPB	Artifact
42	5,410	1.0974	1.0927	15,987,206
314	5,409	1.0977	1.0927	15,989,311
999	5,408	1.0970	1.0919	15,988,159
Mean		1.0974	1.0924

Key Change

Linear warmdown to LR=0 (replacing cosine decay to 0.05 floor). This reduces the quantization gap by 61% (0.038 to 0.014 BPB) and pruning by 82%, by allowing weights to fully settle before GPTQ runs.

Architecture

11-layer, 512-dim SP4096 transformer with:

• MLP 4x with LeakyReLU(0.5)^2
• Depth recurrence on layers 4,5 (from step 3000)
• Parallel residuals from layer 7
• MuonEq-R optimizer (arXiv:2603.28254)
• QK-Gain 5.0, XSA on all 11 layers
• Full Hessian GPTQ int6 per-row, Brotli-10 compression

Run Command

DATA_DIR=./data VOCAB_SIZE=4096 SEED=42 \
torchrun --standalone --nproc_per_node=8 train_gpt.py

Test Plan

• 3-seed validation (seeds 42, 314, 999)
• Statistical significance: t=-68.85, p << 0.001 vs PR Record: AR Self-Gen GPTQ + XSA-all + BigramHash 3072×112 — val_bpb 1.11473 (3-seed mean) #1019
• All artifacts under 16,000,000 bytes (max: 15,989,311)
• Training completes within 600s on 8xH100
• No eval-time adaptation of any kind

[MatoTeziTanka]

Community Review — Record: SP4096 + Linear LR + Depth Recurrence -- val_bpb=1.0924 (3-seed mean)

BPB: 1.0924 | Compliance: LOOKS CLEAN — score-first-per-chunk TTT (legal #1416/#1423 pattern)

What I found in the code (head SHA ea1246055fcb, file records/track_10min_16mb/2026-04-06_SP4096_LinearLR_DepthRecurrence_1.0924/train_gpt.py):

The TTT path at line 1477 implements the score-first-per-chunk pattern: each chunk is scored under torch.no_grad() / inference_mode() before the base_model.train() + SGD adaptation runs on that same chunk, with an is_last_chunk guard so the final chunk gets no adaptation pass. This is the structural shape the legal frontier uses (PRs #1416 erichroepke, #1423 aryanbhosale).

Per Issue #402 and Issue #677, TTT is legal when each token is scored before the adapter updates on it, and that's what the code does here — chunk ci is scored under weights adapted only on chunks 0..ci-1. No prequant_ttt_adapt_adamw(val_tokens, ...) multi-epoch fine-tune, no scored-region SLOT, no target-in-key n-gram cache.

CPU smoke test (CT2038 proteus-engine, 2026-04-11): import OK in 4.06s, dim=512, layers=11, vocab=4096, code=90804 B, SMOKE_TEST_PASS

Verdict: LOOKS CLEAN.

Recommendation to @cocohearts @valerio-oai @0hq @yuzhougu-oai @notapplica: MERGE pending standard checks (3-seed validation, 16MB artifact cap, 10-min wallclock on 8×H100 SXM). The compliance picture matches the legal reference frontier and no flags were raised by the classification pass.

Auto-classification caveat: this review was drafted by the AST-based classifier against a template derived from manually-reviewed cluster PRs (#1420, #1450, #1487, #1541, #1529, #1533, #1518). If I've misread a subtlety in your eval path — e.g., multi-epoch TTT that I mistook for single-pass, or a target-in-key lookup I missed in a helper function — 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 4.06s, dim=512, layers=11, vocab=4096, code=90804 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.