README.md

metadata

pretty_name: CARE-Bench
configs:
  - config_name: model_inputs
    default: true
    data_files:
      - split: public
        path: model_inputs_public.jsonl
  - config_name: cases
    data_files:
      - split: public
        path: cases_public.jsonl
  - config_name: evaluation_prefixes
    data_files:
      - split: public
        path: evaluation_prefixes_public.jsonl
  - config_name: submission_template
    data_files:
      - split: public
        path: submission_template_public.csv

CARE-Bench v1.1 public data

GitHub link to the project

The public release contains 439 cases and 925 labeled prefixes.

Split	Cases	Prefixes
Development	284	609
Validation	93	181
Public Test 1	62	135

The prefix labels are distributed as follows:

Label	Prefixes
Information needed	249
Self-care or monitor	162
Nonurgent care	342
Urgent care	172

The Dataset Viewer includes four configurations:

• model_inputs: identifiers, split information, and model-facing messages
• cases: released source-grounded cases
• evaluation_prefixes: reference labels and evaluation metadata
• submission_template: submission rows for every released prefix

The cases configuration excludes internal review annotations, construction-plan fields, and target-round planning notes.

The public_test_1 split contains 135 scored prefixes.

Controlled-access Test 2 inputs and labels are excluded from the public repository. Input-only data can be requested through the CARE-Bench Test 2 access request form.