anon-caa-neurips/KLINIKEBENCH-Trajectories

KlinikeBench Trajectories

Model-generated trajectories from KlinikeBench (Clinical-Long-Bench): 333 long-horizon clinical consultations in which a clinician agent interviews an LLM-simulated patient and acts through a clinical tool surface, scored by four gates (diagnosis, required tools, parameterised actions, LLM-judged must-ask coverage; outcome_pass = all four pass). Every clinician runs in the Harbor terminus-2 agent, replying with JSON (analysis, plan, commands) whose shell commands call the benchmark's clinic CLI (clinic say "...", clinic call <tool> '<json>').

13,962 trajectories in total. Patients are simulated; no real patient data is included.

Configs

config	rows	what
leaderboard	8,912	27 clinician models × 333 tasks, one rollout each (gemini-3.6-flash: 254 tasks). Patient simulator and must-ask judge gpt-5.4-mini (gpt-5-mini for the findings-gpt5mini run). 10 patient turns, 30 agent steps (+ tool-dependent extension).
step_cap_ablation	1,000	5 clinicians × 50 held-out tasks × agent-step cap {default 30, 45, 60, none (max_steps = 1000000)}.
patient_simulator_grid	4,000	8 clinicians × 10 patient-simulator models × 50 held-out tasks (judge gpt-5.4-mini). run = sim-<patient>--<clinician>.
qwen3_4b_zero_shot	50	Original Qwen3-4B (no fine-tuning) on the 50 held-out tasks.

The sft/ folder holds the SFT training sets built from leaderboard trajectories (chat messages + meta): clb_sft_allmodels_outcome.jsonl.gz (all-gate passes from every model on the 283 training tasks; used for the released SFT model), clb_sft_diagnosis.jsonl.gz (diagnosis-gate passes) and clb_sft_tools.jsonl.gz (tool-gate passes).

Fields

field	description
subset, run, trial_name	config, Harbor job name, trial id
task_id, split	benchmark task; split = train (283 tasks) or test (50 held-out tasks used for SFT/RL evaluation)
task_dataset	Harbor dataset variant the task came from
clinician_model, patient_model, judge_model	models in the three roles
max_turns, max_steps	patient-turn and agent-step budgets for the episode
reward, outcome_pass, diagnosis, required_tools, parameterised_actions, must_ask	verifier scores (0–1)
n_turns, n_steps, diagnosed, sim_ok, judge_ok	episode length, whether a diagnosis was recorded, simulator/judge health
prompt_tokens, completion_tokens, cached_tokens, cost_usd	clinician usage as reported by the provider
started_at, finished_at, exception	timing; exception type if the trial failed (verifier fields are then null)
transcript	JSON string: clinician ↔ patient dialogue ([{role, content}], user = patient)
agent_steps	JSON string: full terminus-2 trajectory (ATIF steps: prompts, model replies, shell commands, terminal observations)
gates	JSON string: per-gate verifier detail (missing tools, parameter violations, per-topic must-ask judge verdicts)
meta	JSON string: episode metadata (budgets, completion, patient-simulator status)

import json
from datasets import load_dataset
ds = load_dataset("anon-caa-neurips/KLINIKEBENCH-Trajectories", "leaderboard", split="train")
ex = ds[0]
dialogue = json.loads(ex["transcript"]); steps = json.loads(ex["agent_steps"])

Notes

• Benchmark version. leaderboard was run on KlinikeBench v1.0, whose task daemon failed to load the clinical database, so lookup tools answered from per-task data. step_cap_ablation, patient_simulator_grid, qwen3_4b_zero_shot and all SFT/RL runs used the fixed build (= Harbor Hub klinikebench/klinikebench v1.1). On the 50 held-out tasks the difference shows up mainly in the required-tools gate, so compare leaderboard with the other configs with care.
• Terminal recordings, tmux panes and per-episode prompt dumps were dropped (the same content is in agent_steps). Local paths and internal endpoints were replaced by placeholders.
• Scripted oracle runs, smoke tests and pilot runs are not included.
• Not a medical device; model outputs can be wrong and must not be used for clinical decisions.

Citation

If you use these trajectories or SFT sets, please cite the KlinikeBench paper (arXiv:2609.38480):

@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}
}