[Non-Record] Hymba-8L: Hybrid SSM + Sliding Window Attention with 32K Context (1.1470 BPB)#1245
[Non-Record] Hymba-8L: Hybrid SSM + Sliding Window Attention with 32K Context (1.1470 BPB)#1245mkenney2 wants to merge 6 commits intoopenai:mainfrom
Conversation
…470 BPB) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
@mkenney2 Heads up — your submission shows 3 valid seeds (1337, 42, 7) but may be getting flagged as incomplete by automated tooling. The issue is your
Quick fix — rename those fields to match the standard schema (see PR #1019 for reference). That should resolve the seed count issue. (Flagged via the Agora) |
- Rename submission_name -> name, results -> seed_results - Add author, github_id, blurb, date fields - Add exact val_loss/val_bpb means and stds - Add artifact_bytes_mean/max, step_avg_ms_mean - Use full precision values from logs Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@MatoTeziTanka Thank you! Super helpful. |
- Type column supports multiple tags per PR (e.g. Neural + TTT) - Filter JS updated: clicking TTT shows all PRs containing TTT - Reclassified TTT submissions as Neural + TTT - Community: resolved issue #6 (mkenney2 schema fix for PR openai#1245) - Community: posted feedback on issue openai#140 for PR openai#1215 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
Thanks for this submission, I'd like to merge this into the non-record leaderboard. Before we merge, could you change the files' location? The PR title marks this as non-record, but the files are currently under:
For it to be an eligible non-record submission, please move it under |
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
@valerio-oai great! I moved the file. Thanks! |
Summary
This submission uses a hybrid architecture combining Mamba SSM with sliding window attention (SWA), which allows us to train at 32x longer context (32,768 tokens) than the standard baseline (1,024 tokens) under the same compute and time constraints. Unlike full attention which scales quadratically, SWA and Mamba both scale linearly, making long-context training feasible within the 10-minute wall-clock budget.
Building on our previous Hymba submission (1.1873 BPB, 7L), this version adds a systematic ablation study across architecture, regularization, quantization, and evaluation strategies, yielding a -0.040 BPB improvement.
Results