Contextual Answers
Contextual Answers is AI21's named task-specific model for questions about supplied material. Its evaluation separates answer accuracy, recognition of missing information and relevance.
Access not confirmed
Unanswerable questions were part of the evaluation
AI21's June 2024 experiment used 1,000 SQuAD 2.0 examples, split equally between answerable and unanswerable questions. This tested whether the model could withhold an answer when evidence was absent.AI21's evaluation of its Contextual Answers model
Practical guidance. This is a useful distinction for document assistants. A system that always produces a fluent answer can still fail its central task.
Extraction before generation
AI21's term-sheet example used Contextual Answers to extract information, then passed that information to Jamba-Instruct for drafting and revision. The two models had different roles in the workflow.AI21's Contextual Answers and Jamba term-sheet workflow
A named model with an older service surface
Practical guidance. The historical model and its evaluations are documented. A current independent API route or public weight release is not confirmed, so this entry does not promise new access.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Access not confirmed |
| Launch date | |
| Model identifiers | AI21 Contextual Answers |
| Purpose | Question answering grounded in supplied contextAI21's evaluation of its Contextual Answers model |
| Public weights | No public checkpoint is confirmedAI21's evaluation of its Contextual Answers model |
Related models
- Jamba2 Mini
AI21 Labs · Jamba2
Hybrid language model with 12 billion active parameters for document questions, extraction and structured text.
Launch date API · Downloadable - Jamba-Instruct
AI21 Labs · Jamba
Historical hosted instruction version of the original Jamba architecture.
Launch date Access not confirmed