Non-record: 11L MLP3.5x LeakyReLU(0.5)^2 + Full SOTA Stack (mean val_bpb=1.1330, 8xH100)#344
Open
aryanbhosale wants to merge 3 commits intoopenai:mainfrom
Open
Non-record: 11L MLP3.5x LeakyReLU(0.5)^2 + Full SOTA Stack (mean val_bpb=1.1330, 8xH100)#344aryanbhosale wants to merge 3 commits intoopenai:mainfrom
aryanbhosale wants to merge 3 commits intoopenai:mainfrom
Conversation
Built on SOTA (10L, int5/int6, BigramHash, SmearGate, SWA) with 75+ automated experiments across Mac MLX and 1xH100 CUDA. Key findings: - NUM_HEADS=4 with head_dim=128: -0.095 BPB relative improvement - Step-based LR schedule: -0.483 BPB vs wallclock-based - BigramHash(16384): -0.025 BPB vs 10240 - MATRIX_LR=0.03: -0.003 BPB Tested on 1xH100 (800 steps, 600s). Post-quant val_bpb: 1.2756 with sliding window eval stride=256. Known issue: artifact is 17.4MB (over 16MB) due to head_dim=128 increasing params. Needs int4/int5 MLP compression to fit budget.
EthanYangTW
added a commit
to EthanYangTW/parameter-golf
that referenced
this pull request
Mar 22, 2026
…le, EMA, Late QAT, TTT Major rewrite targeting top-5 leaderboard: - 11 layers (from 10), BigramHash reduced to 10240 to fit 16MB - XSA (Exclusive Self-Attention) on last 4 layers - Partial RoPE: 16/64 head dims get position encoding - LN Scale: 1/sqrt(layer+1) dampening on deeper layers - EMA (decay=0.997) replaces SWA - Late QAT: STE int6 enabled only in final 4% of training - TTT: 25-epoch SGD on val data post-quantization - FA3 auto-detection with SDPA fallback - Reverted SwiGLU back to relu² (confirmed worse by openai#340, openai#344)
…H100) - 11L 512d 8H/4KV MLP3x LeakyReLU(0.5)^2 - SmearGate + BigramHash(10240) + TrigramHash(4096) - Value Residual + Gated Attention + XSA-all-11 - Partial RoPE(16/64) + int6 GPTQ-lite + zstd-22 - Score-first TTT (3 epochs, per-layer LR, cosine schedule) - Developed via 30-experiment autoresearch on 1xH100 - Artifact: 8.5MB (under 16MB)
…1.1330, 3-seed, 8xH100) - 31.4M params, 11L 512d 8H/4KV MLP3.5x(1792) - LeakyReLU(0.5)^2, SmearGate, BigramHash(10240), TrigramHash(4096) - Value Residual, Gated Attention, XSA-all-11, Partial RoPE(16/64) - Muon lr=0.03, EMA(0.997), Late QAT, int6 GPTQ-lite + zstd-22 - 3-seed: 1.1334/1.1322/1.1334, mean=1.1330, std=0.0007 - Developed via 30-experiment autoresearch on 1xH100
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.
Record: 11L MLP3.5x LeakyReLU(0.5)^2 + Full SOTA Stack
3-seed mean val_bpb: 1.1330 (std=0.0007) on 8xH100 SXM
Architecture (31.4M params)
Training
Quantization
Evaluation
Development
30-experiment autoresearch loop on 1xH100 (~8 hours). Validated on 8xH100 SXM (RunPod).
Feature ablation (1xH100):