The open machine-learning workbench.
One registry for open-weight models, datasets and evaluation runs — versioned, licensed and reproducible. Pull a checkpoint, inspect its card, and see whether it can reach a mainframe transaction core.
from osai import pipeline
# sentence-similarity — a per-record score, the shape pattern 01 can serve
scorer = pipeline("sentence-similarity", model="BAAI/bge-m3")
print(scorer("Claim 4471: duplicate procedure within 30 days"))Trending models
All models →sentence-transformers · Already ONNX · Pattern 01
sentence-transformers · Already ONNX · Pattern 01
transformers · Directly exportable · Pattern 02
transformers · Directly exportable · Pattern 01
transformers · Directly exportable · Pattern 03
transformers · Already ONNX · Pattern 01
Every checkpoint carries a mainframe readiness verdict.
Exportability is a procurement filter, not an afterthought. We screen each repository for a static ONNX graph, recommend an integration pattern for the z/OS core, and publish the copybook contract alongside the weights.
Mainframe program →Datasets, with provenance attached
Every corpus ships a datasheet: collection method, filtering pipeline, license chain and known gaps. Stream splits without downloading terabytes.
Browse datasets