Record: SP2048 + 3-Layer Recurrence + SWA + BigramHash + Legal TTT — val_bpb 1.0955 (3-seed mean)

[bigbag]

Summary

• val_bpb = 1.0955 (3-seed mean, std 0.0004) | ~15.49 MB | 8×H100 SXM
• First SP2048 submission with SWA + BigramHash + 3-layer depth recurrence + legal TTT
• No SLOT, no n-gram cache — fully compliant

3-Seed Results

Seed	Sliding BPB	TTT BPB	Artifact
42	1.0965	1.0952	15,498,155
314	1.0972	1.0960	15,493,880
999	1.0967	1.0954	15,474,490
Mean	1.0968	1.0955	15,488,842

Key Techniques

1. SP2048 Vocabulary — 2048-token SentencePiece BPE
2. 3-Layer Depth Recurrence (layers 3,4,5) — extends PR Record: ParallelResiduals + MiniDepthRecurrence, 1.1063 BPB / 1.8679 nats, -0.0072 vs PR #1179, -0.0143 vs merged SOTA #1204/Record: MuonEq-R + 3-Layer Recurrence + WD=0.095 + MLR=0.022 + All-Int6 — val_bpb 1.0900 (3-seed mean) #1331
3. Stochastic Weight Averaging (from frac=0.75)
4. BigramHash Embeddings (vocab=2048, dim=128)
5. Legal Score-First TTT (SGD, lr=0.002, 3 epochs) — from PR Record: SP4096 + Depth Recurrence + Parallel Residuals + MuonEq-R + Legal TTT — val_bpb 1.0896 (3-seed mean) #1326
6. Parallel Residuals (from layer 7)
7. MuonEq-R + QK-Gain 5.0 + WD=0.095 + MLR=0.022
8. Full GPTQ int6 + Brotli

Credits

PR #1326 @aryanbhosale, PR #1331 @dexhunter, PR #1204 @msisovic, PR #1218 @clarkkev, PR #1260 @dexhunter, PR #1217 @bigbag

Test plan

• 3-seed validation (42, 314, 999)
• All artifacts under 16,000,000 bytes
• No SLOT, no n-gram cache
• Legal score-first TTT

[bigbag]

Closing — PR contains unrelated files. Will resubmit with clean branch.