Skip to content

Non-record: Legal Neural-Only No-TTT Alt (8xH100) val_bpb=1.1576#947

Open
aamodbhatt wants to merge 1 commit intoopenai:mainfrom
aamodbhatt:record-2026-03-27-legal-neural-no-ttt-alt
Open

Non-record: Legal Neural-Only No-TTT Alt (8xH100) val_bpb=1.1576#947
aamodbhatt wants to merge 1 commit intoopenai:mainfrom
aamodbhatt:record-2026-03-27-legal-neural-no-ttt-alt

Conversation

@aamodbhatt
Copy link
Copy Markdown

Summary

  • adds a second compliance-focused neural-only submission using a larger model preset
  • avoids n-gram/two-pass cache blending and keeps evaluation neural-only
  • includes required files (README.md, submission.json, train_gpt.py, train_seed1337.log)

Run Result (seed 1337)

  • val_bpb: 1.15758536 (final_research_export_exact)
  • val_loss: 1.95453440
  • pre-quant diagnostic val_bpb: 1.1399
  • train time: 563.076s
  • eval time: 44.296s
  • total size: 14,921,440 bytes

Compliance Notes

  • NGRAM_EVAL_ENABLED=0
  • NGRAM_TWO_PASS_ENABLED=0
  • NGRAM_FULL_RESCORE=0
  • TTT_ENABLED=0
  • no tokenizer or dataset modifications
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

1 participant