Non-Record: BPB 1.13872 — LeakyReLU(0.5)² + Per-Layer LR Legal TTT (3 seeds)

[Christopher-Lee-McClendon]

Non-Record: 11L LeakyReLU(0.5)² + Per-Layer LR Legal TTT (3 seeds)

val_bpb = 1.13872 (best seed) | Mean: 1.13936 ± 0.0008 | Pre-TTT mean: 1.1574 | Artifact: 15.36 MB

Non-record unlimited-compute submission (trained on 4×A100-40GB, ~42 min; eval ~3690s on 1×A100).

What Changed vs PR #526 (BPB 1.14252)

1. LeakyReLU(0.5)² (from PR Record: 11L XSA4 + LeakyReLU(0.5)² + Cosine TTT 50ep (val_bpb=1.0622) #518, sofiabod): Replace ReLU² → −0.0035 mean pre-TTT improvement. This accounts for essentially all of the final BPB gain.
2. Per-layer LR for TTT (from PR Record: Cosine TTT scheduling with per-layer lr — mean val_bpb=1.0970 (3 seeds) #481, mrdavtan): mlp.proj 3×, mlp.fc 0.5× learning rate
3. Intra-chunk cosine LR (from PR Record: 11L XSA4 + LeakyReLU(0.5)² + Cosine TTT 50ep (val_bpb=1.0622) #518): Cosine decay within each chunk's 30 TTT epochs

Note on TTT modifications: Per-layer LR and intra-chunk cosine were adopted from other PRs but showed no measurable TTT improvement in this data — TTT gain went from −0.0184 (PR #526) to −0.0182 (this PR), a slight regression. The entire final BPB improvement comes from the better pre-TTT model via LeakyReLU. These TTT modifications require further ablation.

3-Seed Validation

Seed	Pre-TTT BPB	Final BPB	Δ vs PR #526
1337	1.1572	1.13912	−0.00340
42	1.1580	1.14024	−0.00228
7	1.1569	1.13872	−0.00380
Mean	1.1574	1.13936	−0.00316

Architecture

• 11L depth-recurrence (10 unique BlockCores), d=512, 8 heads, 4 KV heads
• LeakyReLU(0.5)² MLP, Partial RoPE (16/64), Value Embeddings (128d, layers 9-10)
• SmearGate (input embeddings), BigramHash(2048), XSA(last 4), U-Net skips, LN Scale
• SWA, Late QAT, int6+zstd quantization
• 15,357,245 bytes total (4.0% headroom under 16MB limit)

TTT Protocol (Legal)

• Score-first: torch.inference_mode() during scoring, then train
• SGD(lr=0.002, momentum=0.9), 30 epochs/chunk, freeze first 2 blocks
• Per-layer LR: mlp.proj 3×, mlp.fc 0.5×, intra-chunk cosine decay

Credits

This submission integrates work from many contributors:

• LeakyReLU(0.5)² — PR Record: 11L XSA4 + LeakyReLU(0.5)² + Cosine TTT 50ep (val_bpb=1.0622) #518 (sofiabod)
• Per-layer LR for TTT — PR Record: Cosine TTT scheduling with per-layer lr — mean val_bpb=1.0970 (3 seeds) #481 (mrdavtan)
• Intra-chunk cosine LR — PR Record: 11L XSA4 + LeakyReLU(0.5)² + Cosine TTT 50ep (val_bpb=1.0622) #518 (sofiabod)
• 30-epoch legal TTT + score-first protocol — Our prior work (PRs Non-record: 11L Depth Recurrence + High-Yield Legal TTT (1.14458 BPB) #461, Non-record: 11L + 30-Epoch Legal TTT (BPB 1.14252) #526)
• 11L depth recurrence — PRs Record: 11L Tight SWA + VE128 + XSA4 + TTT (3-seed mean val_bpb=1.1299) #455, Record: 11L EMA + AdamW TTT 10ep (mean val_bpb=1.1027) #442
• Partial RoPE, VE128, LN Scale — PRs Record: 11L + Tight SWA + Shared VE128 + Partial RoPE + LN Scale + XSA4 (val_bpb: 1.1246) #374, Record: 11L Tight SWA + VE128 + XSA4 + TTT (3-seed mean val_bpb=1.1299) #455
• SmearGate, BigramHash, XSA — Community contributions across multiple PRs
• Muon optimizer — PR Record: 11L + Tight SWA + Shared VE128 + Partial RoPE + LN Scale + XSA4 (val_bpb: 1.1246) #374 and descendants
• SWA + Late QAT + int6/zstd — Evolved across many PRs

[MatoTeziTanka]

Community Review — Non-Record: BPB 1.13872 — LeakyReLU(0.5)² + Per-Layer LR Legal TTT (3 seeds)

BPB: 1.13872 | Compliance: LOOKS CLEAN — score-first-per-chunk TTT (legal #1416/#1423 pattern)

What I found in the code (head SHA 3b619c756cd8, file records/track_non_record_16mb/2026-03-23_11L_LeakyReLU_PerLayerLR_LegalTTT/train_gpt.py):

The TTT path at line 849 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 0.05s, dim=512, layers=11, vocab=1024, code=74210 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 0.05s, dim=512, layers=11, vocab=1024, code=74210 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.