NLP Capabilities

Text tasks are first-class in Orinth, not an appendix to the vision workflow. They use the same projects, the same Dataset Studio, the same training and testing surfaces.

Supported tasks

Task	What it produces	Headline metrics
Text classification	A label per record	Accuracy, F1, MCC, AUC
Summarization	A generated summary	ROUGE-1, ROUGE-2, ROUGE-L
Question answering	An extracted or generated answer	Exact match, F1

Preparing text datasets

Text records go through the same lifecycle as images: upload or import, label, split, preprocess, and freeze as a version. The records tab renders text rows for review and relabeling, and the EDA tab reports class balance the same way it does for image classes.

To build a text dataset from documents you already have, see Data Recipes.

Models

NLP training draws on Hugging Face transformers (BERT, BART, and related architectures) and Keras text classifiers. Gated Hub repositories need a HUGGINGFACE_HUB_TOKEN set under Settings.

Testing text models

Per-item inspection puts the original text, the reference answer, and the model's prediction in adjacent columns — which is usually faster at explaining a bad ROUGE score than the score itself.

Where NLP ends and LLM begins

Text classification, summarization, and question answering are supervised tasks with a scored, held-out split. Instruction fine-tuning a language model is a different workflow with a different artifact and a different way of being evaluated — that lives under LLM Fine-tuning & Chat.