Value Residual + DenseFormer DWA + TTT#491
Closed
ahmettrkck wants to merge 6 commits intoopenai:mainfrom
Closed
Conversation
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Replace full-weight TTT with LoRA rank-8 on Q+V+LM head - Per-document adaptation with reset between docs - Multi-epoch (5) with cosine LR decay (0.01 -> 0.001) - Score-then-train per 256-token chunk Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Short docs called forward() without LoRA which returns 0-dim mean scalar, then tried to index it with nll[0]. Use forward_logits + cross_entropy instead. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Start from PR #549 base (33 techniques, 1.1194 bpb) - Add CTW-weighted n-gram backoff (orders 2-7, logistic domain mixing) - Replace heuristic alpha with provably optimal Bayesian model averaging - Keep legal score-first TTT Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Logistic domain mixing was wrong for target-probability mixing. PR #753 uses linear: p_mixed = (1-a)*p_neural + a*p_ngram. Keep CTW-inspired depth-adaptive alpha boost. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
V_used = 0.5 * (V_n + V_1). Zero extra parameters, ~10% parameter efficiency gain (arXiv:2410.17897).Built on top of the current #1 submission (10L_Int5MLP_MuonWD04_SWA50 by thwu1).
Status
Pending 8xH100 run logs. Will update with 3-seed results once compute is available.
Test plan