Skip to content

Notable: SP8192 + 3-Layer Recurrence + Parallel Residuals - 5-Seed Quantization Reference and SDClip Ablations#1720

Open
kiyoaki wants to merge 1 commit intoopenai:mainfrom
kiyoaki:submission/2026-04-19_SP8192_3LayerRecur_ParResid_QK525_NoTTT
Open

Notable: SP8192 + 3-Layer Recurrence + Parallel Residuals - 5-Seed Quantization Reference and SDClip Ablations#1720
kiyoaki wants to merge 1 commit intoopenai:mainfrom
kiyoaki:submission/2026-04-19_SP8192_3LayerRecur_ParResid_QK525_NoTTT

Conversation

@kiyoaki
Copy link
Copy Markdown

@kiyoaki kiyoaki commented Apr 18, 2026

Summary

Test plan

…antization Reference and SDClip Ablations

val_bpb 1.08181 (5-seed mean, std 0.00052) under the 10min_16mb track.
Filed as a notable non-record submission; does not claim a new SOTA
technique. Packages a 5-seed pre/post-quant reference for the current
SP8192 / 3-layer recurrence / parallel residual stack plus unsuccessful
SDClip variants as a documented negative baseline for quantization
tuning under the 16 MB cap.
resouer added a commit to resouer/parameter-golf that referenced this pull request Apr 19, 2026
Round34's openai#1720 reference lane used a packed wrapper whose hidden defaults caused
remote vocab autodetection to fall back to 1024. This patch makes the intended
SP8192 surface explicit without touching the packed payload.

Constraint: Round34 should test openai#1720 as an SP8192 reference, not a launcher-derived sp1024 variant
Rejected: Rely on remote_helper packed-code execution path | it still returns 1024 here
Confidence: high
Scope-risk: narrow
Directive: Any packed PR replay that depends on evaluator autodetect should surface VOCAB_SIZE/DATA_PATH/TOKENIZER_PATH plainly in the wrapper header
Tested: python3 -m py_compile train_gpt.py evaluate.py remote_helper.py; python3 remote_helper.py detect-vocab
Not-tested: remote rerun after relaunch
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