Record Submission: Poly5 Softcap + Z-Loss + YaRN + Zstd-22 + Stride-16 (on PR #549 stack)#1325
Open
monisha-max wants to merge 2 commits intoopenai:mainfrom
Open
Conversation
added 2 commits
April 4, 2026 11:40
…6 (on PR openai#549 stack) Six orthogonal improvements on the current SOTA (PR openai#549, 1.1194 BPB): - Polynomial degree-5 softcap replacing tanh (from ternary PR openai#640) - Z-loss regularization (1e-4 * logsumexp^2) for sharper gradients - YaRN positional encoding for better long-context handling - zstd-22 compression (7MB artifact vs 16MB with LZMA-6) - Sliding eval stride=16 (4x more context overlap) - FlashAttention 3/2/SDPA graceful fallback Smoke-tested on 1xH100 (Modal): 890 steps, healthy convergence. Awaiting 8xH100 SXM verification for official scoring.
1. Adaptive Focal Cross-Entropy Loss - Dynamically upweights hard tokens using model confidence - Focuses limited training budget on informative tokens - (1-p_t)^gamma weighting, normalized to preserve LR scale 2. Residual Vector Quantization (RVQ) - Two-pass: int6 base + int4 residual = ~10-bit effective precision - Exploits 9MB artifact headroom from zstd-22 compression - First application of RVQ to LLM weight compression in this challenge 3. Progressive Depth Warmup - Train layers bottom-up in 3 stages (bottom 1/3 -> 2/3 -> all) - Zero gradients for frozen layer banks after backward - Novel application of gradual unfreezing to from-scratch training
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Six orthogonal improvements on the current SOTA (PR #549, 1.1194 BPB):
1e-4 * logsumexp²) — anchors logits near zero through quantizationConservative estimated gain: -0.007 to -0.014 BPB → target ~1.105-1.112 BPB
Smoke Test (1xH100, Modal)
Status
Smoke-tested on 1xH100 via Modal. Awaiting 8xH100 SXM verification, currently without RunPod/8xH100 access. Happy to collaborate with anyone who can run the 3-seed verification.
All techniques are individually proven in merged PRs (#549, #640, #414). This submission combines them for the first time.
Run Command
Credits
Built on the work of @abaybektursun (PR #549, #399), @signalrush (PR #414), @CiprianFlorin-Ifrim (PR #640), @Christopher-Lee-McClendon (PR #461), @parinzee (PR #493), @jfprincz (PR #315, #287), @unnir (PR #265).
Test plan