Skip to content

Record: GDN-Hybrid + Sliding Window Attention (cold-cache, 1.01710 BPB)#1622

Closed
joshkmartinez wants to merge 1 commit intoopenai:mainfrom
joshkmartinez:submission-run039-safe019-restage
Closed

Record: GDN-Hybrid + Sliding Window Attention (cold-cache, 1.01710 BPB)#1622
joshkmartinez wants to merge 1 commit intoopenai:mainfrom
joshkmartinez:submission-run039-safe019-restage

Conversation

@joshkmartinez
Copy link
Copy Markdown

Summary

  • restages Joshua's strongest still-authoritative clean SAFE_SUBMISSION artifact after the run055-reeval-gdn-bpbfix audit invalidated the later run051-safe031 score
  • stages run039-safe019 (GDN-Hybrid + Sliding Window Attention, cold-cache 3-seed confirmation)
  • uses pulled TensorPool artifacts as authority for the reported metrics

Headline metrics

  • Lane: SAFE_SUBMISSION (fixed-predictor / no-TTT Track-A)
  • 3-seed mean quantized_bpb: 1.01710033
  • 3-seed std: 0.00133490
  • Best seed: 1.016192
  • Artifact size range: 15,522,111 to 15,981,262 bytes
  • Max artifact bytes: 15,981,262 (< 16,000,000)

Per-seed authoritative results

Seed Steps EMA BPB Quantized BPB XSA BPB Artifact bytes
314 2223 1.007670 1.016476 1.020950 15,522,111
777 2239 1.007192 1.016192 1.020919 15,814,260
2718 2240 1.009535 1.018633 1.023874 15,981,262
Mean 1.008132 1.01710033 1.021914 15,772,544.33

Why this restage exists

A later Joshua submission (run051-safe031, previously advertised at 1.01671233 BPB) was re-audited on TensorPool by run055-reeval-gdn-bpbfix. The pulled audit artifacts showed that score used non-canonical SentencePiece byte-accounting; the corrected mean quantized BPB was 1.19671450, not 1.01671233. That makes the later PR stale as a leaderboard claim. This PR restages the strongest still-authoritative Joshua-owned clean artifact instead.

Legality notes

  • fixed int6 model
  • no TTT / no SLOT / no eval-time adaptation in the scored artifact
  • all pulled artifacts stayed below the 16 MB cap
  • submission authority is quantized_bpb from pulled artifacts

Provenance

@joshkmartinez
Copy link
Copy Markdown
Author

Superseded by #1624 after pulled / showed corrected mean quantized val_bpb 1.01710023 with max counted bytes_total 16,060,435 (over the 16,000,000-byte cap). Restaging the strongest currently clean, audited Joshua artifact instead.

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