Record: Order-Adaptive BackoffMixer (mean val_bpb=0.5440)

records/track_10min_16mb/2026-03-26_OrderAdaptive_BackoffMixer/README.md

@@ -0,0 +1,38 @@
+# Order-Adaptive BackoffMixer
+
+## Results
+
+| Seed | val_bpb | Eval time |
+|------|---------|-----------|
+| 42   | 0.5437  | ~391s     |
+| 1337 | 0.5450  | ~391s     |
+| 2024 | 0.5434  | ~391s     |
+| **Mean** | **0.5440** | |
+| Std  | 0.0008  | |
+
+Artifact: ~16.0 MB. Train: 600s on 8xH100 SXM. Eval: ~391s.
+
+## Architecture
+
+11-layer transformer with XSA-all, full MHA (8/8 heads), LeakyReLU(0.5)^2, 3.5x MLP expansion. int5 quantization + compression. EMA, Tight SWA, Soft-Round QAT.
+
+## Eval-Time Scoring
+
+Order-adaptive entropy-gated BackoffNgramMixer with multi-order backoff (2-7 gram). Per-order entropy thresholds for mixing weight selection. Score-first, backward-looking, deterministic.
+
+## Acknowledgments
+
+Standing on the shoulders of many contributors to this competition:
+
+- @abaybektursun — PR #549 (base architecture, Legal TTT framework, Parallel Muon)
+- @deanbrr — PR #659, #779 (original n-gram eval cache concept, BackoffNgramMixer, drift-free TTT)
+- @Asukabot0 — PR #715, #727 (XSA-all, backoff concept, entropy-adaptive alpha)
+- @gowtham0992 — PR #606 (int5 + Soft-Round QAT)
+- @signalrush — PR #414 (EMA training recipe)
+- @sofiabod — PR #518 (LeakyReLU activation)
+- @thwu1 — PR #180 (mixed quantization, BigramHash, SmearGate)
+- @RoyiRa — PR #700 (TTT framework extensions)
+- @Christopher-Lee-McClendon — PR #461 (TTT recipe)
+- @raahilshah — PR #162 (int6 quantization baseline)
+
+This competition has been an incredible collaborative learning experience. Every improvement builds on ideas shared openly by the community.

records/track_10min_16mb/2026-03-26_OrderAdaptive_BackoffMixer/submission.json

@@ -0,0 +1,9 @@
+{
+  "name": "Order-Adaptive BackoffMixer",
+  "val_loss": 0.5440,
+  "bytes_total": 15995959,
+  "blurb": "11L XSA-all 8/8 MHA, order-adaptive entropy gating + BackoffNgramMixer. Mean 0.5440 (std 0.0008) across 3 seeds. Eval 391s.",
+  "author": "hypery11",
+  "github_id": "hypery11",
+  "date": "2026-03-26"
+}