PR #1598

open

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

by amrayachView on GitHub
val_bpb
1.0813
Architecture
Transformer
Optimizer
Artifact Size

Training Techniques

Test-Time Training
score-first TTT
parameters: {"seeds":5}
Other
other
Non-record evidence package documenting canonical D base, R-series sweep results, and negative results under the 16 MB cap
parameters: null
Compression
Brotli
level: null

Novel Contributions

  • Packages a canonical 5-seed D base as reviewer-friendly evidence
  • Includes a best measured single-seed follow-up result for R1_e_baseline
  • Documents that OWC/CDQuant improve raw BPB but incur a fixed-Brotli compression-entropy penalty that breaks the 16 MB cap
  • Provides a self-contained evidence folder with logs, summaries, and artifact map
  • Consolidates the archived seed-0 script and helper chain into a single counted train_gpt.py file