Improved baseline: recurrent depth + LoRA + MTP#21
Closed
monroestephenson wants to merge 7 commits intoopenai:mainfrom
Closed
Improved baseline: recurrent depth + LoRA + MTP#21monroestephenson wants to merge 7 commits intoopenai:mainfrom
monroestephenson wants to merge 7 commits intoopenai:mainfrom
Conversation
8 unique layers x 2 recurrence = 16 effective layers (vs baseline 9), dim=640 (vs 512), SwiGLU MLP, quantization-aware training noise. Estimated 14.9MB artifact under 16MB cap.
- Replace SwiGLU with ReLU^2 (proven at this scale, fewer params per layer) - Remove fake QAT noise (was corrupting weights, not real quantization-aware training) - Replace dead parameterless recurrence_norms with learned per-recurrence scale vectors - Config: 5 unique x 2 rec = 10 effective layers, dim=704, ~9.0MB artifact - Est ~2.1x baseline speed, ~6500 steps, ~3.4B tokens in 10min
- 4 unique x 3 rec = 12 effective layers (was 5x2=10) - dim=768, LoRA rank 32 (was dim=704, rank 16) - Drop fullgraph=True for compatibility with LoRA conditionals - ~6.7K steps estimated in 10 min, 5.72 MB artifact (36% of budget)
Replace enable_gqa kwarg (requires 2.5+) with manual repeat_interleave for KV heads.
Based on baseline architecture (9L/512d) with proven competition wins: - Sliding window eval (stride=256) for ~0.03-0.04 BPB improvement - seq_len=4096 for better per-token context - Tuned Muon optimizer (0.99 momentum, lower LR, longer warmdown) - FP16 tied embedding preservation through quantization - GQA fix for PyTorch 2.4 compatibility - Previous run log saved (1.2844 pre-quant, 2.3258 post-quant) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
sunnypatneedi
pushed a commit
to sunnypatneedi/parameter-golf
that referenced
this pull request
Mar 31, 2026
- logs/daily_research.md: append 2026-03-31 research section - PR openai#771 CLOSED (score-first TTT rule violation) - PR openai#727 CLOSED (n-gram illegal — no renormalization) - Merged SOTA: 1.1147 (PR openai#1019, 2026-03-25) - New PRs: openai#1184 (0.9485 Scylla tokenizer), openai#1185 (0.9641) - SLOT eval technique, Full GPTQ, QK-Gain 4.0 documented - CLAUDE.md: update Competition Strategy + lessons 21-24 - Merged SOTA updated to 1.1147 - Current Best Path rewritten for 2026-03-31 - Lessons openai#21-24: TTT fix, n-gram risk, Scylla, SLOT - TTT constraint clarified to score-first protocol - Version bumped to v9.0 https://claude.ai/code/session_015z6QKyKzDSYzTniW1GPhAe
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
2 participants
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
MODEL_DIM=704with 8 attention heads and 4 KV headsEMA_START_STEPis not reachedConfig: SP-1024, 5 unique x 2 recurrence = 10 effective layers, dim=704, 8 heads, 4 KV heads, tied embeddings.
Test plan
6410127bytes on random weights)0.017730712890625)python3 -m py_compile records/track_10min_16mb/2026-03-18_ImprovedBaseline/train_gpt.py