KlinikeBench (Harbor RL environment)

Paper · Project page · Harbor Hub · SFT & RL code

333 clinician-authored clinical consultations as Harbor tasks. In each task the agent plays the clinician: it interviews an LLM-simulated patient and uses a clinical tool sandbox (chart, labs, imaging, case-specific study results) through the clinic CLI, then records a diagnosis. The verifier scores four gates: diagnosis, required tools, parameterised actions and must-ask coverage (LLM-judged); the benchmark score (pass@1) requires all four.

The task directories here are identical to klinikebench/klinikebench v1.1 on Harbor Hub, laid out for Harbor's git-repos dataset support: registry.json + tasks/<id>/.

Run

Pass the full URL to --repo (a bare org/name defaults to GitHub). The patient simulator and the must-ask judge call an OpenAI-compatible endpoint; the paper's protocol uses gpt-5.4-mini, which is the default.

uv tool install harbor                       # Harbor >= 0.21
export OPENAI_API_KEY=...                    # patient simulator + must-ask judge
harbor run --repo https://huggingface.co/datasets/anon-caa-neurips/klinikebench \
  -a terminus-2 -m <provider/model>

harbor datasets list --repo https://huggingface.co/datasets/anon-caa-neurips/klinikebench   # list only

Run a single task with -i <task_id> (repeatable; globs allowed). Override the patient / judge with CAA_PATIENT_BASE_URL, CAA_PATIENT_MODEL, CAA_PATIENT_API_KEY and CAA_JUDGE_*.

RL training

Every task's verifier writes /logs/verifier/reward.json with all gates. The scalar reward is chosen with CAA_REWARD_MODE:

mode	reward
outcome (default)	1 if all four gates pass (the benchmark score)
multigate	(2·diagnosis + required_tools + parameterised_actions + must_ask) / 5
diagnosis	diagnosis gate only

The paper trains Qwen3-4B with SFT and then GRPO (SkyRL + Harbor) on 283 training tasks and evaluates on 50 held-out tasks; the code, the train / held-out task-id lists and the released checkpoints are in Zehui127/klinikebench-train. CAA_MAX_STEPS_OVERRIDE sets an absolute agent-step cap (or none) for budget ablations.

API keys. Harbor passes [environment.env] into the task container, where the agent's shell can read it. For training or leaderboard runs, put the key behind a host-side proxy and point CAA_PATIENT_BASE_URL at it (see the training repository).

Layout

registry.json                 Harbor registry: one dataset, klinikebench@1.1 (333 tasks)
tasks/<task_id>/
  task.toml  instruction.md
  environment/  Dockerfile  entrypoint.sh  runtime/  task.json  clinical_db.json  [findings.json]
  tests/  test.sh  verify.py  caa_gates.py  caa_judge.py
  solution/solve.sh           scripted oracle (scores 1.0)

Patients are simulated by an LLM; no real patient data is included. Research use only — not a medical device.

Citation

@misc{fang2026klinikebench,
  title         = {KlinikeBench: Evaluating Language Models Beyond Diagnostic Accuracy},
  author        = {Fang, Xueting and Li, Zehui and Yang, Yang and Giovino, Camilla and Patel, Shubh K. and
                   Prajapati, Shailly and Subasri, Vallijah and Shan, Caihua},
  year          = {2026},
  eprint        = {2609.38480},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2609.38480}
}