atma-10b-babilong-2k-ft-nope

ATMA checkpoint adapted on BABILong qa1 through qa10 using only the 0K, 1K, and 2K configurations. This is an internal controlled adaptation checkpoint, not an official BABILong leaderboard submission.

Protocol

  • Base model: ChavyvAkvar/atma-10b-L40S-mbs16-nope__reg-baseline__distr-0__mem-1__win-0
  • Base revision: d2974b46bf1be80fe0285ed7c15903b8079f615a
  • Dataset revision: ee0d588794c7ac098062ee0d247c733d62e94fe2
  • Prompt protocol: builtin-v1
  • Training lengths: 0k, 1k, 2k
  • Sequence length: 2048
  • Train rows: [0, 80]
  • Validation rows: [80, 90]
  • Reserved test rows: [90, 100]
  • 256K pilot status: complete
  • Full 0K through 256K evaluation status: complete

Load this repository with the checkpoint-exact ATMA benchmark code. It is not a Transformers AutoModel checkpoint.

Downloads last month
4
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ChavyvAkvar/atma-10b-babilong-2k-ft-nope

Dataset used to train ChavyvAkvar/atma-10b-babilong-2k-ft-nope

Collection including ChavyvAkvar/atma-10b-babilong-2k-ft-nope