Skip to content

Non-record: SP8192 D 5-seed base and R-series evidence package#1598

Open
amrayach wants to merge 1 commit intoopenai:mainfrom
amrayach:submission/sp8192-d-rseries-non-record
Open

Non-record: SP8192 D 5-seed base and R-series evidence package#1598
amrayach wants to merge 1 commit intoopenai:mainfrom
amrayach:submission/sp8192-d-rseries-non-record

Conversation

@amrayach
Copy link
Copy Markdown

@amrayach amrayach commented Apr 13, 2026

Summary

This PR adds a non-record submission package for the SP8192 D base and the 2026-04-09 R-series sweep.

The package is intended as a self-contained evidence folder under records/track_non_record_16mb/ and does not claim a new SOTA or a submission-valid lead result.

Key points

  • Canonical base: D (SP8192 #1413 family with parallel residual + loop adjustment)
  • Canonical result: 1.08128837 score-first TTT BPB (5-seed mean, sigma = 0.00058943)
  • Best measured single-seed follow-up: R1_e_baseline = 1.08078562
  • Important caveat: R1_e_baseline ran at 605s, so it is not a clean lead submission number
  • Main measured finding: on this stack, OWC/CDQuant improve raw BPB but create a fixed-Brotli compression-entropy penalty that breaks the 16 MB cap

What is included

  • README.md
  • REPORT.md
  • submission.json
  • requirements.txt
  • single-file package-local train_gpt.py derived from the archived seed-0 script and helper chain
  • canonical D train logs for seeds 0, 42, 1234, 1337, 2025
  • d_submission_summary.tsv
  • r1_e_baseline.log
  • r_series_combined_summary.tsv
  • ARTIFACT_MAP.md

Important framing

  • This is a non-record submission.
  • The package’s main evidence claim is anchored to the completed RunPod D 5-seed bundle.
  • Pegasus rerun failures are documented as operational context only, not as a missing primary-evidence requirement for the package.
  • The package-local train_gpt.py is a single counted code file derived from the archived seed-0 runtime chain for submission review.

Notes

  • Relative to the submission base, this branch only adds one new folder under records/track_non_record_16mb/.
  • The package includes a minimal requirements.txt, but reproducing the measured runs still assumes the matching CUDA/Hopper environment with flash_attn_interface available.

@amrayach amrayach force-pushed the submission/sp8192-d-rseries-non-record branch from bd4e023 to 5d8cdba Compare April 13, 2026 15:10
@amrayach amrayach changed the title Non-record: SP8192 D/R-series evidence package Non-record: SP8192 D 5-seed base and R-series evidence package Apr 13, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant