Record: sliding eval, FP16 tied embeddings, 10 layers, Muon WD 0.02, overtone init, and phase-transition residual mixing. (val_bpb 1.1876)

records/track_10min_16mb/2026-03-20_TopRecipe10L_MuonWD_OvertoneInit_PeytonTolbert/README.md

@@ -0,0 +1,31 @@
+# Top Recipe 10L + Muon WD + Overtone Init
+
+Single-seed 8xH100 record-track run using the upgraded trainer configured to match the strongest known recipe family as closely as possible:
+
+1. Sliding-window final evaluation with `stride=64`
+2. FP16 tied embedding export
+3. 10 transformer layers
+4. Muon weight decay `0.02`
+5. Overtone spectral embedding initialization with power `0.5`
+6. Phase-transition residual-mix initialization
+
+## Result
+
+- Pre-quant eval at stop: `val_loss:2.0602`, `val_bpb:1.2202`
+- Post-quant eval: `val_loss:2.0053`, `val_bpb:1.1876`
+- Exact printed metric: `final_int8_zlib_roundtrip_exact val_bpb:1.18762449`
+
+## Run Details
+
+- Hardware: `8x H100 80GB`
+- Train wallclock: `599846ms`
+- Steps completed: `8260`
+- Final eval time: `118402ms`
+- Total submission size int8+zlib: `15842628` bytes
+- Code size: `128619` bytes
+
+## Included Files
+
+- `train_gpt.py`: exact trainer used for the run, copied from the upgraded trainer
+- `train.log`: full training and evaluation log
+- `submission.json`: leaderboard metadata

records/track_10min_16mb/2026-03-20_TopRecipe10L_MuonWD_OvertoneInit_PeytonTolbert/submission.json

@@ -0,0 +1,16 @@
+{
+  "author": "Peyton Tolbert",
+  "github_id": "peytontolbert",
+  "name": "Top Recipe 10L + Muon WD + Overtone Init",
+  "blurb": "8xH100 single-seed run matching the strongest current recipe family: sliding eval, FP16 tied embeddings, 10 layers, Muon WD 0.02, overtone init, and phase-transition residual mixing.",
+  "date": "2026-03-20",
+  "val_loss": 2.00525163,
+  "val_bpb": 1.18762449,
+  "pre_quant_val_bpb": 1.2202,
+  "pre_quant_val_loss": 2.0602,
+  "steps": 8260,
+  "ms_per_step": 72.62,
+  "eval_time_ms": 118402,
+  "bytes_total": 15842628,
+  "bytes_code": 128619
+}