Record: 1.0400 BPB -- Hedge Mixer + VRL + AdamW TTT + Polyak EMA

[pentxayc]

Summary

• 1.0400 BPB (seed 42, 2 additional seeds pending)
• 11L transformer (26.99M params) with Value Residual Learning (VRL), LeakyReLU(0.5)², XSA-4
• 5-expert Hedge Mixer during eval: neural model + unigram + bigram + trigram (64K hashed) + entropy
• Hedge algorithm (eta=0.1) with deferred between-chunk weight updates (legal score-first)
• AdamW TTT (lr=0.0005) + Polyak EMA (decay=0.998) + byte-weighted loss + adaptive cosine LR
• Freeze first 9/11 blocks during TTT, unfreeze last 2 + norms/scales
• Int6 mixed quantization + lzma compression
• Artifact: 15,999,919 bytes (under 16MB limit)
• Training: 6104 steps in 600s on 8xH100 SXM
• Eval (TTT + Hedge): 404s / 600s budget

Legality

All eval-time adaptations are strictly score-first:

1. Hedge weights for chunk N computed from chunks 0..N-1 only (deferred update after all windows scored)
2. N-gram tables updated after chunk scoring completes
3. Polyak EMA uses fixed decay, no snapshot selection
4. TTT trains only on already-scored chunks
5. No validation data during training; no training data during evaluation

Test plan

• Seed 42: 1.0400 BPB
• Seed 1337: pending
• Seed 2024: pending

🤖 Generated with Claude Code

[MatoTeziTanka]

Community Review — Record: 1.0400 BPB -- Hedge Mixer + VRL + AdamW TTT + Polyak EMA

BPB: 1.0400 | Compliance: LOOKS CLEAN — score-first-per-chunk TTT (legal #1413 dexhunter pattern)

What I found in the code (head SHA 6cff4df0d716, file records/track_10min_16mb/2026-03-25_HedgeMixer_VRL_AdamWTTT_1.0400/train_gpt.py):

The TTT path at line 1017 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 of the current leaderboard's legal frontier (PR #1413 dexhunter, the 1.0828 SP8192 + QK-Gain 5 + Legal TTT entry — verified at its head SHA against the is_last_chunk + torch.no_grad() score-first accumulator pattern).

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 0.12s, dim=512, layers=11, vocab=1024, code=94305 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 0.12s, dim=512, layers=11, vocab=1024, code=94305 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.