Non-record: Paid Prefix Research (val_bpb=1.0539, ruled out-of-scope)#275
Closed
ibarrajo wants to merge 1 commit intoopenai:mainfrom
Closed
Non-record: Paid Prefix Research (val_bpb=1.0539, ruled out-of-scope)#275ibarrajo wants to merge 1 commit intoopenai:mainfrom
ibarrajo wants to merge 1 commit intoopenai:mainfrom
Conversation
Hybrid compression approach: 8L SmearGate/Int6 model (11.67MB) + LZMA-compressed val tokens (4.24MB, 10% coverage) = 1.0539 BPB. Approach banned by organizers but submitted as research contribution exploring the compression-vs-modeling tradeoff at the heart of this competition. Key finding: prefix coverage matters more than model quality in this regime. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Collaborator
|
Guys, come on. |
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
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
track_non_record_16mb/as research contributionKey Finding
Prefix coverage matters more than model quality. Our strong 8L model + small prefix (1.0539) was outperformed by PR #168's weak 7L model + large prefix (1.0238). Each MB of prefix removes more BPB than each MB of model in this regime.
Optimal (unexplored): 3L model (~3MB) + bigram-rank encoded prefix (~13MB, ~46% coverage) → estimated ~0.75 BPB.
Why Submit This
Even though organizers ruled val-token storage out-of-scope, this work explores the fundamental question: what is this competition measuring? BPB is a compression metric. The line between "model that compresses" and "direct compression" is a design choice. This submission documents that boundary, plus practical compression research (LZMA vs pack10 vs bigram-rank encoding).
Contents
README.md— Full writeup with compression analysis and budget tradeoff datatrain_gpt.py— PR 11-Layer Int6 + WD=0.04 + SWA + FA3 (val_bpb: 1.1318) #198 rebase with SDPA fallback + paid prefix supporttrain.log— 8xH100 training log (8L + 7M-token prefix variant)submission.json— MetadataTest plan
🤖 Generated with Claude Code