RoBERTa QA - Extractive Question Answering on SageMaker

deepset/roberta-base-squad2 is RoBERTa-base fine-tuned on SQuAD 2.0, achieving F1 82.9 and EM 79.9 on the dev set. As an extractive QA model, it returns answers as literal character spans from the input document -- if the answer is not in the document, it returns no answer rather than generating one. This structural property makes it suitable for regulated industries where hallucination is not acceptable.

Deploy as a SageMaker endpoint for document QA, policy lookup, contract interrogation, and any workflow where answers must be grounded in a specific source document. All processing stays in your AWS VPC -- no document content leaves your account.

The endpoint accepts {"question": "...", "context": "..."} and returns {"answer": "...", "score": 0.0-1.0, "start": int, "end": int}. The start/end character offsets allow the calling application to highlight the exact source span in the original document.

Note: This model is trained on English. License is CC-BY-4.0 -- attribution to deepset is required in your product documentation and listing.

Primary use cases: regulated-industry document QA where auditability is required, policy and procedure lookup, contract clause interrogation, technical documentation search, and any RAG pipeline where the answer must be traceable to a source span.