PROTEUS v9 — 11L INT6 + single-epoch LoRA TTT (mean val_bpb=1.1526, 3 seeds)#633
Open
MatoTeziTanka wants to merge 1 commit intoopenai:mainfrom
Open
PROTEUS v9 — 11L INT6 + single-epoch LoRA TTT (mean val_bpb=1.1526, 3 seeds)#633MatoTeziTanka wants to merge 1 commit intoopenai:mainfrom
MatoTeziTanka wants to merge 1 commit intoopenai:mainfrom
Conversation
… seeds) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Author
|
Adding nats (val_loss) for clarity:
Platform: RunPod 8×H100 SXM, PyTorch 2.8.0+cu128. Docker: |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Result
Mean val_bpb: 1.1526 (3 seeds, std: 0.0004)
TTT Legality — Single Epoch, Score-Then-Train
This submission addresses the ruling on PR #568 (comment) where @valerio-oai correctly identified multi-epoch TTT as training on eval data.
This submission uses
TTT_EPOCHS=1. Each token is scored exactly once, before being trained on:No token is ever scored after being trained on. This is the same score-then-train pattern as PR #77 (merged), applied once per document.
What changed from v7/v8:
TTT_EPOCHSreduced from 2-5 to 1.TTT_MIN_DOC_LENset to 512. No other code changes.Architecture
11 transformer layers, 512d, 8/4 heads (GQA), 26.8M params. INT6 GPTQ-lite quantization, zstd-22 compression. Artifact ~15.4 MB (96.3% of 16MB budget).
Full details in README.md.
Platform
RunPod 8×H100 SXM, PyTorch 2.8.0+cu128.