Record: MTP-2 Funnel + LeakyReLU(0.75)² + Legal TTT + Parallel Muon

[michaelwinczuk]

Summary

• Added Multi-Token Prediction (MTP_NUM_HEADS=2, MTP_LOSS_WEIGHT=0.1) as auxiliary training signal
• MTP forces backbone to learn richer representations by predicting 2 tokens ahead
• MTP heads discarded at export — zero 16MB impact, zero eval overhead
• Validated -0.0037 BPB improvement on test pod (apples-to-apples, same hardware)
• Lighter weight (0.1 vs default 0.2) avoids gradient stealing from main CE loss

Changes from prior submission (val_bpb 1.1185)

• MTP_NUM_HEADS: 0 → 2
• MTP_LOSS_WEIGHT: 0.2 → 0.1

Changes from SOTA baseline

• negative_slope: 0.5 → 0.75
• MATRIX_LR: 0.025 → 0.027
• WARMDOWN_ITERS: 3500 → 3700
• MTP_NUM_HEADS: 0 → 2
• MTP_LOSS_WEIGHT: 0.2 → 0.1

Research methodology

8 swarm missions + external cross-validation identified MTP as highest-ROI unexplored lever. The "training funnel" concept: MTP auxiliary loss focuses gradient signal on structurally important tokens without adding parameters to the final checkpoint.

🤖 Generated with Claude Code

[MatoTeziTanka]

Community Review — Record: MTP-2 Funnel + LeakyReLU(0.75)² + Legal TTT + Parallel Muon

BPB: 0.0037 (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 02eb7a8d204b, file records/track_10min_16mb/2026-03-27_LeakyReLU075_LegalTTT_ParallelMuon_TunedLR/train_gpt.py):

The TTT path at line 1074 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 0.03s, dim=512, layers=11, vocab=1024, code=89459 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 0.03s, dim=512, layers=11, vocab=1024, code=89459 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.