11L MLP2x + LeakyReLU² + Legal TTT (val_bpb=1.2201, 3-seed mean, std=0.0015)#1057
Open
Programmerryoki wants to merge 1 commit intoopenai:mainfrom
Open
11L MLP2x + LeakyReLU² + Legal TTT (val_bpb=1.2201, 3-seed mean, std=0.0015)#1057Programmerryoki wants to merge 1 commit intoopenai:mainfrom
Programmerryoki wants to merge 1 commit intoopenai:mainfrom
Conversation
…3-seed mean, std=0.0015) Records folder with full 3-seed validation. Config: 11L/512d, 8H/4KV GQA, MLP 2x, seq_len 2048 Training: Muon + AdamW, EMA(0.997), int6 QAT + zstd-22, 600s on 8xH100 Eval: Sliding window stride=64 + Legal score-first TTT (7ep SGD) 3-seed results: Seed 1337: val_bpb=1.2184, 7821 steps, 14,986,599 bytes Seed 42: val_bpb=1.2212, 7833 steps, 15,010,826 bytes Seed 2025: val_bpb=1.2207, 7863 steps, 14,980,707 bytes Mean: val_bpb=1.2201 (std 0.0015) All artifacts under 16MB.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
3-Seed Results
Configuration
11L / 512d / 8H / 4KV (GQA), MLP 2× with LeakyReLU(0.5)², seq_len 2048, BigramHash(4096), SmearGate, U-Net skips, XSA (last 4), LN Scale 1/√(L+1), EMA(0.997), int6 QAT + GPTQ-lite + zstd-22
Eval Protocol
Sliding window eval (stride=64) + legal score-first TTT: SGD(lr=0.002, momentum=0.9), 7 epochs per already-scored 32K chunk, cosine LR decay, all blocks unfrozen. Same protocol as PR #549.
Attribution
Built on PR #414 stack, LeakyReLU² (PR #493), Legal TTT (PR #461), XSA (PR #198), and community contributions.
Full details in records/track_10min_16mb/2026-03-29_11L_MLP2x_LeakyReLU_TTT/README.md