Skip to content

Add Nuclear Stack submission: 1.16668 BPB (seed 2884431328)#178

Closed
timowhite88 wants to merge 3 commits intoopenai:mainfrom
timowhite88:submission/NuclearStack_FarnsworthTech
Closed

Add Nuclear Stack submission: 1.16668 BPB (seed 2884431328)#178
timowhite88 wants to merge 3 commits intoopenai:mainfrom
timowhite88:submission/NuclearStack_FarnsworthTech

Conversation

@timowhite88
Copy link

Int6 + 3x MLP + SmearGate + BigramHash + SWA + TTT with honest eval. More seeds running — will be added as they complete.

Int6 + 3x MLP + SmearGate + BigramHash + SWA + TTT with honest eval.
More seeds running — will be added as they complete.
Seeds: 1337 (1.16516), 2884431328 (1.16668). Honest eval, no double-counting.
@timowhite88
Copy link
Author

@0hq Ready for review — 2-seed mean 1.16592 BPB, third seed comin when credits allow

Key points:

  • Every rule met: 15.8MB artifact (under 16MB), 600s training, 341s eval (under 600s), 8xH100 SXM
  • Honest eval: We fixed the sliding-window double-counting bug that inflates other submissions' scores. Each validation token is scored exactly once
  • Full reproducibility: Complete training logs for both seeds included, single command to reproduce
  • Two orthogonal improvements stacked: architectural (int6, 3x MLP, SmearGate, BigramHash, SWA) + test-time training — no other submission combines both

3-seed mean: 1.16759 BPB
Seed 2884431328: 1.16668 BPB
Seed 1337: 1.16516 BPB
Seed 7: 1.17091 BPB
@timowhite88
Copy link
Author

3-seed submission complete! Added seed 7 (val_bpb=1.17091).

3-seed results:

Seed val_bpb val_loss Steps ms/step
1337 1.16516 1.96733 7248 83.06
2884431328 1.16668 1.96988 7009 85.60
7 1.17091 1.97704 6466 92.79

Mean BPB: 1.16759 | Best: 1.16516 (seed 1337)
Artifact: 15.8MB (int6+zstd) | All runs 8xH100 SXM

@0hq Ready for review!

@timowhite88
Copy link
Author

Superseded by #254 (FarnsworthEngine v1 — 1.1303 BPB with 3-seed validation). Closing this one.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant