ClinSeekAgent_RL

RL-ready prompt sets derived from Letian2003/DeepMed_trajectory.

The source release ships agent rollout trajectories: every question appears 4–5 times, once per sampled run, each carrying the full messages transcript and that run's outcome. That shape is built for SFT and trajectory analysis. This derivative strips it back to what an RL loop actually needs — the prompt and its ground-truth label, once per question — so the policy generates its own trajectories instead of replaying recorded ones.

Every task here is agentic retrieval: the prompt names a patient by subject_id and states the question, and the policy is expected to pull the relevant record itself through EHR tools. No subset hands the model a pre-assembled patient record.

Processing

1. Dropped rollout columns — messages, status, completed, stop_reason, error, prediction_source, score, predictions. These describe one recorded run's behaviour and outcome, which an on-policy learner supplies itself.
2. Dropped run-identity columns — run_index, session_id, runs_per_question. After deduplication these carry no information.
3. Deduplicated by qid, preferring a healthy run. Runs that errored out leave a degraded row — a failed MCP session, for instance, sets error and drops multimodal_summary — so taking whichever run came first would have kept a damaged row while a clean sibling sat right behind it. Selecting on run health rescued 909 qids in multimodal_trajectory_opus47 and 4 in EHR_Bench_trajectory_opus47_clean. Only 2 qids (both multimodal) have no healthy run at all, and they are missing multimodal_summary.
4. Dropped input from EHR_Bench_trajectory_opus47_clean — upstream, that column carried the patient's whole record serialized into the prompt (median 16.8k characters), making the task long-context reading comprehension rather than retrieval. Removing it aligns the subset with the other two and keeps MIMIC record text out of this repo. It also cuts the file from 182 MB to 43 MB.

Everything else passes through unchanged — no reformatting, no renaming. Ground truth stays in label, and task metadata, patient identifiers, and multimodal file pointers are preserved as-is.

Which source files are here

The upstream release has five files; three are kept.

EHR_Bench_trajectory_opus47_clean is kept and the larger EHR_Bench_trajectory_opus47 is not. The _clean file is a strict subset of it — all 6,779 of its qids appear in the larger file, and every shared row is byte-identical — so shipping both would duplicate 6,779 prompts. Worth knowing what the upstream _clean filter removed: the 1,264 dropped questions are not rollout failures (both groups are ~99.8% status: success). They skew toward much longer serialized records (median 29.3k vs 16.8k characters), toward answer leakage (the gold answer appeared verbatim in the upstream input column for 59% of dropped rows against 33% of kept ones), and toward the high-volume ICU event-stream tasks — inputevents loses 258 of 295, procedureevents 228 of 299, outputevents 156, datetimeevents 154.

results_with_answers is not included.

The three kept subsets have no overlap with one another: no shared qids, no shared question text, and disjoint task types.

Subsets

AgentEHR_train3k_3ep

• Rows: 3,000 (one per unique qid)
• Tasks: diagnoses_ccs (500), labevents (500), microbiologyevents (500), prescriptions (500), procedures_ccs (500), transfers (500)
• Columns: qid, question, subject_id, hadm_id, prediction_time, task, label

EHR_Bench_trajectory_opus47_clean

• Rows: 6,779 (one per unique qid)
• Tasks: diagnoses_ccs (299), diagnoses_icd (299), admissions (298), diagnosis_ccs (298), next_event (298), poe (298), procedures_ccs (298), radiology (298), labevents (296), omr (296), services (295), emar (294)
• Columns: qid, question, subject_id, hadm_id, prediction_time, latest_event_time, task, task_type, label, source_idx, source_file, instruction, output, candidates, task_info

multimodal_trajectory_opus47

• Rows: 10,000 (one per unique qid)
• Tasks: ehrxqa_table (3844), medmod_length_of_stay (2451), medmod_decompensation (2418), ehrxqa_image (1117), medmod_radiology (93), ehrxqa_image_table (39), medmod_phenotyping (26), medmod_in_hospital_mortality (12)
• Columns: qid, question, multimodal_summary, subject_id, hadm_id, stay_id, prediction_time, task, task_family, scope, label, answer_type, source_benchmark, source_split, db_path_hint, modalities, study_ids, dicom_ids, image_paths, report_paths, snapshot_spec, source_current_time, official_split, _mm_has_image, _mm_has_ehr_subject

Caveats

• You need your own credentialed EHR data. These rows reference MIMIC-IV and MIMIC-CXR by subject_id, hadm_id, image_paths, and db_path_hint; they do not contain patient records. Both resources are credentialed through PhysioNet under its Credentialed Health Data Use Agreement.
• prediction_time is the snapshot cutoff for the patient record — evidence after it should not be visible to the agent. EHR_Bench_trajectory_opus47_clean also carries latest_event_time, the timestamp of the last event before that cutoff.
• Paths under multimodal_summary.resolved_image_paths are absolute paths from the original authors' machine; use the relative image_paths / report_paths.
• In EHR_Bench_trajectory_opus47_clean, output restates the gold answer already in label, and task_info restates label, task, source_idx, and the two timestamps. Both are kept for fidelity to the source, not because they add information. task_info.metric names the metric the benchmark scores with (em, recall, acc), which is useful when writing a reward function.

License

Apache-2.0 covers this derivative's processing and packaging. It does not and cannot relicense the underlying clinical data, which remains subject to the PhysioNet Credentialed Health Data Use Agreement.

Citation

Please cite the original authors — this repo is a reformatting of Letian2003/DeepMed_trajectory, not new data collection.