NoCap-Test evidence · think2earn 3/3 · ≈8.3%
Source code ↗

BottleCapAI NoCap-Test

≈8.3% less train time to val ≤ 3.3821 — 3/3 seeds

Carol Calin · ML Engineer and NLP Researcher, BottleCapAI NoCap-Test · acalincarol@gmail.com

NoCap-Test trains GPT-2 ≈124M on FineWeb on one GPU until validation loss reaches ≤ 3.3821, faster than a fair stock baseline. The score is first-passage train time: clock time until the run first crosses that threshold.

The challenge is about algorithmic ideas that might generalize, not hyperparameter tuning or copying Modded-NanoGPT wholesale.

We adopt Keller Jordan’s Muon (explainer; published Ortho/NS, not our invention). Plain Muon alone failed or lost under our seal (failure ledger). What we sealed is Muon on a B32 systems path plus timing-cheap split-QKV after a costlier split was killed (what QKV is; kill → bank → 3/3). Joint-QKV Muon on the same stack missed the quality bar where measured. Speed is vs our sealed AdamW F1_seal (3.866 h, explained here): 3/3 seeds, mean ≈8.3%. Not an official NoCap win, and not a matched % vs joint Muon stock. Sealed trainer + SHA freeze: TRAINER_DELTA · kit/ (this site) · view source ↗.

What the number means

Three seeds first-cross val_disjoint ≤ 3.3821 in 3.51–3.57 h (mean 3.544 h, mean ΔT% ≈ +8.3% vs F1_seal 3.866 h). Terminal candidate CE ≈ 3.37. Per-seed T and ΔT% are in the scoreboard and seeds explainer.

How it is measured

3 / 3
seeds under F1_seal
≈8.3%
mean ΔT% vs F1
3.866 h
F1_seal T_cross
≈3.37
candidate terminal CE

Pin: 1× A100-SXM4-40GB · torch 2.13.0+cu130 · GPT-2 ≈124M · FineWeb · official disjoint val. Workspace codes: Clock-W, Claim-W1, Claim-W2.

Not claimed. Not an official NoCap win — only under our sealed F1. Matched candidate-vs-stock first-cross ΔT% is N/A (joint-QKV stock never crossed on 1337/2029; seed 4242 stock not run). Public % vs F1_seal only. Why each clause.