Record: XSA-all + Depth Recurrence + Hedge Mixer TTT (val_bpb=1.0278, 3-seed mean)#733
Closed
stukenov wants to merge 2 commits intoopenai:mainfrom
Closed
Record: XSA-all + Depth Recurrence + Hedge Mixer TTT (val_bpb=1.0278, 3-seed mean)#733stukenov wants to merge 2 commits intoopenai:mainfrom
stukenov wants to merge 2 commits intoopenai:mainfrom
Conversation
…(val_bpb=1.0278, 3-seed mean) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…ucibility Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Author
|
Closing: eval time exceeds 600s limit. Resubmitting with TTT_EPOCHS=1. |
pappanick
added a commit
to pappanick/parameter-golf
that referenced
this pull request
Mar 26, 2026
- Per-head learned gate in attention (PR openai#638/openai#733): -0.002 BPB - Lambda_v * x0 shortcut from initial embedding (PR openai#657/openai#733): -0.002 BPB - Both enabled by default via GATED_ATTENTION=1, VALUE_RESIDUAL=1 - Added attn_gate, lambda_v to control tensor patterns for proper quantization handling - All smoke tests pass on CPU
pappanick
added a commit
to pappanick/parameter-golf
that referenced
this pull request
Mar 26, 2026
…eader Major additions: - Depth recurrence: layers 4,5 repeated -> 13 virtual from 11 physical Repeat blocks share heavy CastedLinear weights, own scalar params untie_recurrence() deep-copies before TTT for independent specialization Only ~1% param overhead during training - TTT defaults changed to match PR openai#733 winning recipe: - SGD optimizer (was AdamW) - simpler, less memory - lr=0.002 (was 0.0005) - higher for SGD - Unfreeze all 11 blocks (was 2) - more params for adaptation - All repeat_blocks params unfrozen for TTT Configurable via: RECUR_LAYERS="4,5" TTT_OPTIMIZER=sgd TTT_LR=0.002 All smoke tests pass on CPU (syntax, recurrence, weight sharing, untie).
7 tasks
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.
Record: XSA-all + VRL + CROWN-Q + Depth Recurrence + Hedge Mixer TTT
val_bpb = 1.0278 (3-seed mean, std 0.0039) | ~15.8 MB | 8xH100 SXM, 600s train
3-Seed Results
Key Innovations (6 additions over PR #549)
Legal TTT (Score-First)
Every token scored under
torch.inference_mode()BEFORE any weight update. Hedge Mixer n-gram tables built from already-scored tokens only. SGD optimizer (not AdamW) for TTT.Note on eval time
TTT eval takes ~755s (exceeds 600s limit). Reducing
TTT_EPOCHSfrom 3 to 1 brings eval under 600s with expected BPB ~1.08-1.09. Happy to resubmit with 1 epoch if required.Reproduction
All defaults in the script match the submitted results. No env vars needed.
Credits
PR #549 (@abaybektursun), PR #634 (@raahilshah), PR #657 (@anthony-maio), PR #638 (@Asukabot0), PR #693 (@EthanYangTW), PR #686 (@msisovic), PR #688 (@RoyiRa), PR #493 (@parinzee), PR #414 (@signalrush)