Record: MuonEq-R + 3-Layer Recurrence + WD=0.095 + MLR=0.022 + All-Int6 — val_bpb 1.0900 (3-seed mean)#1331
Open
dexhunter wants to merge 1 commit intoopenai:mainfrom
Open
Conversation
…b 1.0900 (3-seed mean) 3-layer depth recurrence (layers 3,4,5) with WD-LR synergy: higher WD (0.095) compresses for all-int6 headroom, higher MLR (0.022) recovers quality. All 66 layers at int6 precision. 3-seed mean: 1.0900 BPP / 2.5077 nats (seeds 42, 0, 7) All seeds under 16MB with 36K+ margins. No TTT, no SLOT, no eval-time adaptation. Improves PR openai#1285 (1.0912) by 0.0013 BPP. Beats PR openai#1218 by 0.0079. Built on PR openai#1218 by @clarkkev.
This was referenced Apr 4, 2026
This was referenced Apr 7, 2026
6 tasks
manfromnowhere143
added a commit
to manfromnowhere143/parameter-golf
that referenced
this pull request
Apr 18, 2026
Two complementary additions to AwebUltimate base: 1. Depth Recurrence (PR openai#1517/openai#1331/openai#1471 pattern, must credit): Re-run selected encoder layers once more after encoder pass, with NO skip-stack push (preserves U-Net 5-in/5-out symmetry). Curriculum: activated at RECUR_START_STEP (default 2000). Eval always uses recurrence. Env vars: RECUR_LAYERS (e.g. '3,4'), RECUR_START_STEP. 2. Lookahead Optimizer (Zhang/Lucas/Hinton/Ba, NeurIPS 2019) — Aweb signature: Maintains slow weights for all trainable params. Every k inner steps: slow := (1-α)*slow + α*fast; fast := slow. ~5%% wall-clock overhead. Novel for nanochat speedrun (verified via gh search). Env vars: LOOKAHEAD_ENABLED, LOOKAHEAD_K, LOOKAHEAD_ALPHA. Backwards-compat: RECUR_LAYERS='' (default) + LOOKAHEAD_ENABLED=0 reproduces proven 1.1190 baseline byte-identically. CPU smoke test (10 cases) PASSES: env wiring, model construction, recur parsing, forward/forward_hidden recur application, skip-stack symmetry, mini-training loss decrease, lookahead update math.
G3sparky
added a commit
to G3sparky/parameter-golf
that referenced
this pull request
Apr 18, 2026
QK_GAIN_INIT=5.5 extends the monotonic improvement trend past 5.25. 3-seed mean 1.0809 (std 0.0004) on 8xH100 SXM. Base: SP8192 + 3-Layer Depth Recurrence + Parallel Residuals + Legal TTT (PRs openai#1394, openai#1331, openai#1437, openai#1412, openai#549, openai#1445)
5 tasks
6 tasks
5 tasks
This was referenced Apr 26, 2026
8 tasks
This was referenced Apr 30, 2026
Closed
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
Key Innovation: 3-Layer Recurrence + WD-LR Synergy
Extends 2-layer recurrence (PR #1285) to 3 layers. The extra virtual layer needs more artifact budget, compensated by:
Results
Changes from PR #1285 (1.0912)
Credits