Non-record: TTT-LoRA Base — HumanAI Convention (val_bpb=1.2364)#600
Open
humanaiconvention wants to merge 1 commit intoopenai:mainfrom
Open
Non-record: TTT-LoRA Base — HumanAI Convention (val_bpb=1.2364)#600humanaiconvention wants to merge 1 commit intoopenai:mainfrom
humanaiconvention wants to merge 1 commit intoopenai:mainfrom
Conversation
Non-record submission introducing per-document Test-Time Training via rank-128 LoRA adapters. Base score val_bpb=1.23637747 with TTT disabled; TTT LR calibration in progress for follow-up competitive submission. Architecture: 11L dim=512 GQA SmearGate OrthoInit SWA int6+zstd22 Novel: TTT LoRA adaptation during 10-min eval budget (orthogonal to all current leaderboard entries) Co-Authored-By: Claude Sonnet 4.6 <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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Submission: TTT-LoRA Base
Track:
track_non_record_16mbAuthor: HumanAI Convention (@humanaiconvention)
val_bpb: 1.23637747
Submission size: 15,669,326 bytes
Novel contribution
This submission introduces per-document Test-Time Training via LoRA as an evaluation-time technique. During the 10-minute evaluation budget, each validation document receives its own fresh rank-128 LoRA adapters trained with Adam on the preceding chunks before predicting the next chunk.
This approach is orthogonal to all current leaderboard entries and exploits the separate evaluation budget in a way no current submission does.
Base submission note
TTT is disabled (
TTT_ENABLED=0) in this submission. Local testing showed consistent improvement (−0.136 bpb on 50-doc smoke test), but the first production run revealed that the LR tuned on smoke models (0.001) is too high for a well-trained production model. LR calibration is in progress. A follow-up submission with working TTT targeting competitive val_bpb is planned pending compute credits.Architecture changes vs baseline
Compute credits request
We are applying for OpenAI compute credits via the grant form to complete TTT LR calibration.