Non-record: BESE + Mamba-3 SSD Hybrid (1.3571 BPB, 7.6 MB artifact)#1665
Open
mrbese wants to merge 2 commits intoopenai:mainfrom
Open
Non-record: BESE + Mamba-3 SSD Hybrid (1.3571 BPB, 7.6 MB artifact)#1665mrbese wants to merge 2 commits intoopenai:mainfrom
mrbese wants to merge 2 commits intoopenai:mainfrom
Conversation
added 2 commits
April 16, 2026 00:56
First combination of a custom byte-level tokenizer (BESE, 288 vocab) with Mamba-3 SSD hybrid architecture (6 Mamba + 2 Attention). Addresses the "State-space models" bounty from the challenge README. val_bpb: 1.3571 (INT6 + LZMA + sliding window + n-gram tilt) Artifact: 7,614,888 bytes (48% of 16 MB limit)
…folder reproduction, code-structure note - Add full inline BPB correctness proof (per-token byte accounting + invariant + per-case argument + transitive merge accounting + runnable self-test) - Add explicit single-seed disclaimer for the 1.3571 headline; clarify that train_log_run2/run3 are architecture ablations, not seed replicates - Add 'Quick path' reproduction that runs train_gpt.py directly from the records folder (per the rule that submissions must run within the records folder) - Keep full-pipeline reproduction instructions for the data-prep step - Add 'Code Structure' section addressing the FAQ rule about train_gpt.py
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Architecture
8 layers: 6 Mamba-3 SSD blocks + 2 Attention blocks at positions [2, 5]. dim=512, d_state=128, ngroups=1, expand=2. No depth recurrence (hurts SSMs per PR #1355). Pure PyTorch SSD implementation.
Key findings
Ongoing work
Compute credits pending for: Triton kernel integration (2-3x faster steps), TTT, QAT for wider models, and 3-seed statistical runs. Target: 1.17-1.20 BPB.
Files
README.md— Full writeup with architecture details, results, ablations, and reproduction stepssubmission.json— Metadatatrain_gpt.py+mamba3_ssd.py+ tokenizer files — Self-contained, runnable from records folder