The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
Kefir2 finite training qualification
Diagnostic only. This is NOT the production stratified training dataset.
Two fixed blocks of 64 sequences, 2049 positions per row (2048 next-token objectives),
from public abcastor/kefir2-casebits, region 1. They reuse the original global-planner
code on a single-leaf subset; this subset cannot establish production mixture or mixing quality.
Original NVIDIA Nemotron-ClimbMix attribution and CC-BY-NC-4.0 restrictions apply.
CaseBits mixture base IDs and complete ASCII case masks are separate uint16 arrays.
valid is a boolean useful-target mask: first position is context, padding remains
in the historical optimizer/LR objective and is separately counted. EOD base ID is0.
Do not silently mask padding out of the objective or change memory boundary behavior.
The prefix contains exact source receipt/file hashes, split rule and plan/batch identities.
A finite two-block cursor must stop at exhaustion; never wrap it into a long training run.
No credentials, model weights or private conversations are included. Every member is
SHA256-bound by manifest.json. Load arrays with allow_pickle=False.
- Downloads last month
- 30