Available on AWS Marketplace · Version 1.0.0
Sources, licenses and distribution
The offline runtime includes intfloat/multilingual-e5-small at revision
614241f622f53c4eeff9890bdc4f31cfecc418b3, under the MIT license. Its original
license notice is included in the image. Software Sushi's original serving
code and frozen ridge head use Apache-2.0. The encoder is not represented as
an original Software Sushi pretrained model.
The upstream model card is available from
the pinned publisher revision.
The publisher's pretraining history is not a claim of project-owned customer data.
Dependency licenses, Python/native package metadata and an unused Lingua dependency
notice remain in the container. No external business-document dataset or customer
record has been admitted. No training corpus is shipped in the runtime.
All 8,331 project examples are fictional/model-authored, including dependent
variants. The head fits 3,924 TRAIN rows through 459 family means. Exposed DEV
influenced head selection. A separate frozen 324-row generated final was not fitted;
its provisional labels and AI-review disagreements remain disclosed in the model
card. Replaying these exposed rows on AWS checks portability, not new accuracy.
There are zero native documents and zero independent human acceptance labels.
Generated examples and AI reviewers do not establish customer accuracy or human
adjudication. No relabeling/exclusions were used to remove final disagreements.
The seven-type assisted scope excludes OCR, mixed/unknown detection, authenticity,
field extraction and automatic decisions. Out-of-scope text still gets suggestions.
A person must choose or reject both suggestions. Validate with your own authorized
representative documents before operational use. Model rank is not confidence.