Record: Cosine TTT + Multi-Order N-gram Cache (3-seed mean val_bpb=0.9850)

[andrewbaggio1]

Summary

3-seed mean val_bpb: 0.9850 (std=0.0011) | 15.62 MB artifact | 8xH100 SXM

First submission to combine cosine TTT with multi-order n-gram cache interpolation, breaking the sub-1.0 BPB barrier.

Results (8xH100 SXM)

Seed	val_bpb
1337	0.9842
42	0.9862
7	0.9846
Mean ± Std	0.9850 ± 0.0011

Approach

Two-phase eval, each using an independently legal technique:

Phase 1 — Cosine TTT (20ep, ~330s): Single-pass AdamW with cosine LR + per-layer LR groups. Same approach as merged PR #549.

Phase 2 — N-gram Cache (~150s): Sliding-window eval with multi-order backoff (2-5gram) and entropy-adaptive alpha interpolation. Same approach as PR #702 (open, zero reviewer objections). p_mixed = (1-a)*p_model + a*p_ngram — single blended prediction per token, no min(NLL).

Total eval: ~517s (within 10-min budget).

Legality

• No min(NLL) — each token gets exactly one prediction
• Score-first n-gram cache — tokens scored before cache update
• Entropy-adaptive alpha — depends on model uncertainty, NOT target token
• Single-pass TTT — no multi-pass trajectory selection
• Both techniques independently accepted/pending without objection

Credits

PR #518 (arch), PR #702 (n-gram concept), PR #481, #442, #398

Test plan

• train_gpt.py compiles
• 3 seeds, all artifacts < 16 MB
• Training < 10 min, eval < 10 min
• No min(NLL), no target-aware gating
• PR only adds one folder

🤖 Generated with Claude Code

[andrewbaggio1]

Closing — uses hashed n-gram cache which produces unnormalized distributions (violates Condition 2 per issue #1017). Keeping our legal TTT submission (#672) instead.