Record: 11L Parallel Muon + N-gram Backoff Cache — val_bpb 0.2841 (3-seed mean)

[aryanbhosale]

Record: 11L Parallel Muon + N-gram Backoff Cache

val_bpb = 0.2841 (3-seed mean, std 0.0001) | ~15.92 MB | 8×H100 SXM

3-Seed Results (8×H100 80GB SXM, PyTorch 2.9.1+cu128)

Seed	step_avg	steps	EMA bpb	Quantized bpb	N-gram bpb
1337	88.6ms	6,774	1.1193	1.1279	0.2841
42	88.6ms	6,772	1.1194	1.1276	0.2840
2024	88.7ms	6,769	1.1191	1.1279	0.2840
Mean	88.6ms	6,772	1.1193	1.1278	0.2841

Key Innovation: N-gram Backoff Cache (Eval-Time Only)

Order 2-9 backward-looking N-gram cache with entropy-adaptive alpha blending and 65K-token chunk updates:

• For each scored token, blend model P(token) with N-gram frequency P(token)
• Alpha adapts based on model entropy + n-gram order (high entropy + high order = more n-gram weight)
• Per-order multipliers: orders 2-3 suppressed (0.3x), orders 5-9 boosted (2.0x)
• Cache updated ONLY after scoring each 65K-token chunk (strictly backward-looking)
• 4M hash buckets, XOR-of-products hashing

N-gram reduces BPB by 4x (1.1278 -> 0.2841) by exploiting repeated phrases and patterns.

Architecture (26.8M params)

11L 512d, 8H/4KV (GQA), MLP 3x LeakyReLU(0.5)², Parallel Muon (parameter banking + batched NS5), SmearGate, BigramHash(1024), Value Residual, Gated Attention, XSA4, Partial RoPE(16/64), EMA(0.997)+SWA, Late QAT, GPTQ-lite int6+zstd-22, FA3, torch.compile(fullgraph=True).

Timing

• Training: 600s (6,772 steps at 88.6ms/step)
• Eval (N-gram): ~420s
• Total: within 600s train + 600s eval budgets

Compliance

• Training under 600s on 8xH100
• Eval under 600s on 8xH100 (~420s)
• Total artifact under 16,000,000 bytes
• N-gram cache strictly backward-looking (updated AFTER scoring)
• No training data access during evaluation
• No oracle/hindsight selection
• 3-seed results with full logs

Credits

• N-gram cache concept: PR Record: 5-gram Eval Cache + LeakyReLU² + Parallel Muon val_bpb: 1.0920 (3-seed mean, std 0.0007) | ~15.9 MB | 8×H100 SXM #659 by @deanbrr, PR Podracing: 1.0461 BPB (3-seed mean) #674 by @newjordan
• Multi-order backoff + entropy-adaptive: PR Record: 1.0240 BPB — Multi-Order N-gram Backoff + Entropy-Adaptive Alpha (100% autonomous research via goldfish) #702 by @lukacf
• Fine-grained chunk updates: PR Record: Order-12 N-gram Backoff + 256K Chunks — 0.2834 BPB #843 by @quietsmile
• Parallel Muon / Parameter Banking: PR Record: Parallel Muon + Parameter Banking — 81.87ms/step, val_bpb 1.1247 (3-seed mean) #399 by @abaybektursun
• LeakyReLU²: PR Record: 11L EMA + Int6 + XSA + LeakyReLU² + Partial RoPE (val_bpb: 1.1309) #493 by @parinzee
• Base model stack: PR Record: 11L EMA + GPTQ-lite + warmdown3500 + QAT@0.15 (val_bpb=1.1233) #414 by @signalrush

[MatoTeziTanka]

Community Review — Record: 11L Parallel Muon + N-gram Backoff Cache — val_bpb 0.2841 (3-seed mean)

BPB: 0.2841 | Compliance: LOOKS CLEAN — score-first-per-chunk TTT (legal #1416/#1423 pattern)

What I found in the code (head SHA 74c599514de6, file records/track_10min_16mb/2026-03-26_11L_ParallelMuon_NgramBackoff/train_gpt.py):

The TTT path at line 749 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 0.09s, dim=512, layers=11, vocab=1024, code=99731 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 0.09s, dim=512, layers=11, vocab=1024, code=99731 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.