UltraBERT Curation

Use it to score or classify text such as web pages, prompts, maths problems, and library records for curation attributes including topic, format, educational value, and difficulty.

Availability

Available on Hugging Face; usage instructions and model files are provided

Pricing model

Not established

Deployment

Runs on CPU or GPU; can be served locally

Privacy

Not established

Licence

Apache-2.0

• The model returns probabilities for options in choice, yes/no (noul), and ordered-level (score) questions. ^[1]
• It is intended for curation questions about web pages, prompts, maths problems, and library records. ^[1]

  | Good at | curation questions on web pages, prompts, maths problems and library records: topic, format, educational value, difficulty, task type, validity, catalogue class |

• The model, training data, and training recipe are released publicly. ^[1]

  A 380M-parameter decision model for data curation, with **its training data and the recipe that made it, both

• The Jev / System One API can run on CPU or GPU. ^[1]

  In Python, without the server: `from jev_api import system_one; system_one(state, questions)`.

• The model was tested in English only, and its API refuses inputs over 4,096 tokens. ^[1]

  - **Length.** The API refuses inputs over 4,096 tokens (the training length); the fast path truncates at 2,048

• The release describes the model as a curation specialist, not a general decision model. ^[1]

  A specialist, not a general decision model |