Record: 11L VR + GA + LeakyReLU² + Legal Score-First TTT (val_bpb=pending)

[amaljithkuttamath]

Summary

val_bpb = pending rerun | 8xH100 SXM | Legal score-first TTT

Architecture improvements on the standard 11L stack: Value Residual + Gated Attention + LeakyReLU(0.5)², with legal score-first TTT.

Update (Mar 23)

Initial submission used pre-eval TTT (non-compliant per #402). This update switches to legal score-first TTT: score each validation chunk before training on it. Score pending rerun with compute credits.

Architecture

• Value Residual (arXiv:2410.17897): Learned lambda blends first-block V into subsequent blocks. +2 params/layer.
• Gated Attention (arXiv:2505.06708): Per-head sigmoid gate on attention output. +4104 params/layer.
• LeakyReLU(0.5)²: Replaces relu² in MLP. Preserves negative gradient flow.
• Legal score-first TTT: AdamW lr=1e-4, 3 epochs, cosine LR across 131K-token chunks, last 2 blocks + norms unfrozen. Every token scored BEFORE model trains on it.

VR+GA ablated in PR #413 (-0.017 bpb combined). LeakyReLU² proven by PR #569 and #535.

Run command

torchrun --standalone --nproc_per_node=8 train_gpt.py

All hyperparameters set as defaults.

Credits

• PR Record: Train Larger, Quantize Harder - 33.6M params + int5 GPTQ / (val_bpb: 1.1164) #576 (cmcdnd): Legal score-first TTT, temperature calibration
• PR Record: 11L VRL + LeakyReLU² + Full GPTQ (3-seed mean val_bpb=1.1175) #569 (gowtham0992): VRL + LeakyReLU² + Full GPTQ
• PR Non-record: Value Residual (-0.015 BPB) + Gated Attention (-0.003 BPB) with ablations #413: VR + GA ablation
• PR Record: 11L Partial RoPE + LN Scale + EMA + XSA4 (val_bpb: 1.1248) #315 (jfprincz): Foundation architecture

[MatoTeziTanka]

Community Review — Record: 11L VR + GA + LeakyReLU² + Legal Score-First TTT (val_bpb=pending)

BPB: 0.017 (cache parse — may be delta/std, not val_bpb; check PR title) | Compliance: LOOKS CLEAN — score-first-per-chunk TTT (legal #1416/#1423 pattern)

What I found in the code (head SHA 3c4c4bd3dfec, file records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/train_gpt.py):

The TTT path at line 1042 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 0.04s, dim=512, layers=11, vocab=1024, code=84868 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 0.04s, dim=512, layers=11, vocab=1024, code=84868 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.