Non-record: SWA and doc-isolated eval ablation — two negative findings at stride=64#199
Open
mrdavtan wants to merge 1 commit intoopenai:mainfrom
Open
Non-record: SWA and doc-isolated eval ablation — two negative findings at stride=64#199mrdavtan wants to merge 1 commit intoopenai:mainfrom
mrdavtan wants to merge 1 commit intoopenai:mainfrom
Conversation
Three controlled runs on fast 8xH100 hardware (44ms/step, 13,600+ steps): - SWA does not improve int8 quantization under default warmdown (+0.0004 BPB) - Doc-isolated eval HURTS by 0.009 BPB at stride=64, contradicting LoRA TTT entry's +0.011 finding at stride=256 - Crossover: context quantity dominates at stride=64, quality at stride=256+ - Includes SWA bf16 accumulation bug discovery and fix
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Three controlled runs on fast 8×H100 hardware (44ms/step, 13,600+ steps) testing SWA and doc-isolated evaluation independently. Both techniques are being adopted by top entries without isolated ablation data.
Finding 1: SWA does not improve int8 quantization (default warmdown)
SWA may help with aggressive warmdown (WD=20000) where the LR stays higher and produces more diverse snapshots. Under default warmdown (WD=1200), the model is already in a narrow basin — averaging nearby snapshots doesn't find a flatter minimum.
Finding 2: Doc-isolated eval HURTS by 0.009 BPB at stride=64
This contradicts the LoRA TTT entry (#77) which found +0.011 from doc-isolation at stride=256. The difference: at stride=64, tokens have 960+ tokens of context. Removing cross-doc context at document boundaries means start-of-document tokens lose ALL context (window starts fresh), which hurts more than cleaner context helps.
The crossover: At stride=64 context quantity dominates; at stride=256+ context quality dominates. If you use stride=64, do not use doc-isolated eval.
Bonus: SWA bf16 accumulation bug
An initial implementation accumulated SWA weights in bf16 over thousands of steps, producing val_bpb=2.62 (catastrophic). Fix: accumulate in float32, sample every 50 steps. Documented for anyone implementing SWA in bf16 pipelines.
Reproduction
Three runs, same seed (1337), same architecture (9L×512d baseline), one variable per comparison. Full commands and analysis in README.
Test plan
Built with Claude Code