Initial upload: 1800 ehr_bench + 989 multimodal mm_bench inputs (matches hierarchical eval qids)

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+ehr_bench/all_inputs.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,284 @@
+---
+license: other
+language:
+- en
+tags:
+- clinical
+- ehr
+- medical
+- multimodal
+- benchmark
+pretty_name: ClinSeek-Bench
+size_categories:
+- 1K<n<10K
+configs:
+- config_name: ehr_bench
+  data_files: "ehr_bench/all_inputs.jsonl"
+- config_name: mm_bench
+  data_files: "mm_bench/all_inputs.jsonl"
+---
+
+# ClinSeek-Bench
+
+Inputs (questions, patient context, gold labels) for the **ClinSeek** clinical-reasoning evaluation reported in
+`UCSC-VLAA/ClinSeek-Evaluation-Results` under `hierarchical/`.
+
+This dataset contains **the exact data points evaluated** — every `qid` in this dataset has matching scored outputs in *both*
+`hierarchical/agentic/` and `hierarchical/oneshot/` of the results dataset. The two modes share an identical question set, so
+the same data point can be evaluated under either an **agentic** rollout (multi-turn tool use) or a **one-shot** rollout
+(single LLM call), producing two scores per data point per model.
+
+## Contents
+
+| Split | # data points | Modality | Source |
+|---|---:|---|---|
+| `ehr_bench` | **1,800** | text-only | MIMIC-IV electronic health records |
+| `mm_bench` | **989** | multimodal (text + chest X-ray) | MIMIC-CXR images linked to MIMIC-IV / EHR-XQA-style questions |
+| **Total** | **2,789** | | |
+
+`mm_bench` here is restricted to **image-grounded** questions only. The text-only multimodal queries (`L2.ehr_query`, n = 1,704
+in the upstream `combined_test_set.jsonl`) were **excluded** because they are EHR-only and not part of the multimodal
+evaluation track.
+
+## Directory layout
+
+```
+ClinSeek-Bench/
+├── README.md                    # this file
+├── manifest.jsonl               # qid → taxonomy + result-path templates
+│
+├── ehr_bench/
+│   ├── all_inputs.jsonl                   # flat: 1,800 rows, all ehr_bench inputs
+│   └── L1.text_only/
+│       ├── L2.decision_making/            # 27 leaves × 40 rows = 1,080 rows
+│       │   └── L3.<sub-family>/
+│       │       └── L4.<task>/
+│       │           └── inputs.jsonl
+│       └── L2.risk_prediction/            # 18 leaves × 40 rows = 720 rows
+│           └── L3.<sub-family>/
+│               └── L4.<task>/
+│                   └── inputs.jsonl
+│
+└── mm_bench/
+    ├── all_inputs.jsonl                   # flat: 989 rows, all multimodal inputs
+    └── L1.multimodal/
+        ├── L2.cxr_vqa/                    # 619 rows: presence / enumeration / change
+        │   └── L3.<sub-family>/L4.<task>/inputs.jsonl
+        ├── L2.risk_prediction_mm/         # 250 rows: 24-h + full-stay mortality
+        │   └── L3.<sub-family>/L4.<task>/inputs.jsonl
+        └── L2.phenotyping_mm/             # 120 rows: CCS phenotype classification
+            └── L3.<sub-family>/L4.<task>/inputs.jsonl
+```
+
+The `L1 / L2 / L3 / L4` tags are an evaluation taxonomy:
+- **L1** — modality (`text_only` or `multimodal`)
+- **L2** — task family (e.g. `decision_making`, `risk_prediction`, `cxr_vqa`)
+- **L3** — sub-family (e.g. `Diagnosis_coding`, `ICU_Mortality`, `cxr_vqa_presence`)
+- **L4** — concrete task (e.g. `diagnoses_ccs`, `ICU_Mortality_3day`, `cxr_finding_presence`)
+
+The taxonomy is identical to the one used by the result repository, so the `(bench, L1, L2, L3, L4)` tuple acts as a stable
+join key across input ↔ result.
+
+## Schema
+
+### `ehr_bench/**/inputs.jsonl`
+
+Each row is one EHR-Bench data point.
+
+```jsonc
+{
+  "idx": 0,                                         // original index in upstream ehr_bench_*.jsonl
+  "qid": "ehr_bench_decision_making_0",             // stable id; matches result rows.jsonl
+  "instruction": "Given the sequence of events ...", // task instruction
+  "input": "## Patient Demographics ...",           // patient context (markdown of structured EHR)
+  "output": "OBSERVATION ADMIT",                    // gold answer (string for single-label tasks)
+  "candidates": ["...", "..."],                     // discrete answer choices (when applicable)
+  "task_info": {
+    "target_key": "admission_type",
+    "metric": "em",
+    "task_type": "decision_making",
+    "task": "admissions",
+    "event": "admissions",
+    "target": "[\"OBSERVATION ADMIT\"]",
+    "label": ["OBSERVATION ADMIT"]                  // gold answer as list (canonical form)
+  },
+  "taxonomy": {                                     // added by us — routing key
+    "bench": "ehr_bench",
+    "L1": "L1.text_only",
+    "L2": "L2.decision_making",
+    "L3": "L3.Transfers_Services_Admissions",
+    "L4": "L4.admissions"
+  }
+}
+```
+
+### `mm_bench/**/inputs.jsonl`
+
+Each row is one multimodal data point with one or more linked CXR images.
+
+```jsonc
+{
+  "qid": "medmod_radiology_test_56445431",          // stable id; matches result rows.jsonl
+  "question": "<task_instruction>...<question>...", // self-contained prompt text
+  "image_paths": ["mimic-cxr/2.0.0/files/p17/p17795062/s56445431/<dicom>.jpg"],
+  "subject_id": 17795062,
+  "study_ids": [56445431],
+  "dicom_ids": ["a69ab575-cc7ec66c-..."],
+  "stay_id": 39825171,
+  "hadm_id": 22792685,
+  "prediction_time": "2177-07-25 22:04:23",
+  "label": [{"name": "Cardiomegaly"}, {"name": "Support Devices"}],   // gold answer
+  "task": "medmod_radiology",
+  "db_path_hint": "database/patient_17795062.db",
+  "source_benchmark": "medmod",
+  "source_split": "test",
+  "task_family": "qa",
+  "answer_type": "list",
+  "taxonomy": {
+    "bench": "mm_bench",
+    "L1": "L1.multimodal",
+    "L2": "L2.cxr_vqa",
+    "L3": "L3.cxr_vqa_presence",
+    "L4": "L4.cxr_finding_presence"
+  }
+}
+```
+
+> **Images are not bundled with this dataset.** The `image_paths` follow the upstream
+> [MIMIC-CXR JPG](https://physionet.org/content/mimic-cxr-jpg/) layout. To reproduce the multimodal evaluation, download
+> MIMIC-CXR-JPG from PhysioNet (it requires credentialed access) and resolve `image_paths` against your local mirror.
+
+### `manifest.jsonl`
+
+A flat join key for the whole dataset. One row per data point:
+
+```jsonc
+{
+  "qid": "ehr_bench_decision_making_0",
+  "bench": "ehr_bench",
+  "L1": "L1.text_only",
+  "L2": "L2.decision_making",
+  "L3": "L3.Transfers_Services_Admissions",
+  "L4": "L4.admissions",
+  "inputs_path": "ehr_bench/L1.text_only/L2.decision_making/L3.Transfers_Services_Admissions/L4.admissions/inputs.jsonl",
+  "result_path_template": "hierarchical/{mode}/ehr_bench/{model}/L1.text_only/L2.decision_making/L3.Transfers_Services_Admissions/L4.admissions/rows.jsonl"
+}
+```
+
+`{mode} ∈ {agentic, oneshot}` and `{model}` is one of the model slugs evaluated in the result dataset
+(e.g. `claude_opus_4_6`, `kimi_k2_5`, `qwen3_5_35b_a3b`, …). The format inside that `rows.jsonl` is:
+
+```jsonc
+{ "qid": "...", "f1": 1.0, "precision": 1.0, "recall": 1.0,
+  "gold": ["..."], "pred": ["..."], "bench": "ehr_bench" }
+```
+
+## Same data point under two modes
+
+Every `qid` here appears **once in `agentic/` and once in `oneshot/`** in the result dataset. To pull both rollouts of one
+data point under one model, use the `result_path_template` from `manifest.jsonl` and substitute `{mode}` twice:
+
+```python
+import json
+
+# Pick one data point.
+with open("manifest.jsonl") as f:
+    rec = json.loads(next(f))               # qid = ehr_bench_decision_making_0
+
+# Load its question + gold from the inputs side (this dataset).
+with open(rec["inputs_path"]) as f:
+    qrow = next(r for l in f if (r := json.loads(l))["qid"] == rec["qid"])
+
+# Load both modes from the result dataset (UCSC-VLAA/ClinSeek-Evaluation-Results).
+def load_score(model, mode):
+    path = rec["result_path_template"].format(mode=mode, model=model)
+    with open(path) as f:
+        return next(r for l in f if (r := json.loads(l))["qid"] == rec["qid"])
+
+agentic = load_score("claude_opus_4_6", "agentic")
+oneshot = load_score("claude_opus_4_6", "oneshot")
+print(qrow["instruction"], "→ gold:", agentic["gold"])
+print("agentic F1:", agentic["f1"], "| oneshot F1:", oneshot["f1"])
+```
+
+The two scores are directly comparable: same prompt, same gold, different agent stack.
+
+## Tasks (L4 leaf table)
+
+| bench | L1 (modality) | L2 (family) | L3 (sub-family) | L4 (task) | N |
+|---|---|---|---|---|---|
+| ehr_bench | L1.text_only | L2.decision_making | L3.Diagnosis_coding | L4.diagnoses_ccs | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Diagnosis_coding | L4.diagnoses_icd | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Diagnosis_coding | L4.diagnosis | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Diagnosis_coding | L4.diagnosis_ccs | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.ICU_Events | L4.chartevents | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.ICU_Events | L4.datetimeevents | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.ICU_Events | L4.ingredientevents | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.ICU_Events | L4.inputevents | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.ICU_Events | L4.outputevents | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Lab_Microbiology_orders | L4.labevents | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Lab_Microbiology_orders | L4.microbiologyevents | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Medication_suggestion | L4.emar | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Medication_suggestion | L4.medrecon | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Medication_suggestion | L4.medrecon_atc | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Medication_suggestion | L4.prescriptions | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Medication_suggestion | L4.prescriptions_atc | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Medication_suggestion | L4.pyxis | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Next_Event | L4.next_event | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Outpatient_Records | L4.omr | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Procedure_coding | L4.procedureevents | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Procedure_coding | L4.procedures_ccs | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Procedure_coding | L4.procedures_icd | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Provider_Orders | L4.poe | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Radiology_Orders | L4.radiology | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Transfers_Services_Admissions | L4.admissions | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Transfers_Services_Admissions | L4.services | 40 |
+| ehr_bench | L1.text_only | L2.decision_making | L3.Transfers_Services_Admissions | L4.transfers | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.ED_Critical_Outcomes | L4.ED_Critical_Outcomes | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.ED_Hospitalization | L4.ED_Hospitalization | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.ED_ICU_Transfer_12h | L4.ED_ICU_Tranfer_12hour | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.ED_Reattendance_3day | L4.ED_Reattendance_3day | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.ICU_Mortality | L4.ICU_Mortality_14day | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.ICU_Mortality | L4.ICU_Mortality_1day | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.ICU_Mortality | L4.ICU_Mortality_2day | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.ICU_Mortality | L4.ICU_Mortality_3day | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.ICU_Mortality | L4.ICU_Mortality_7day | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.ICU_Readmission | L4.ICU_Readmission | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.ICU_StayLength | L4.ICU_Stay_14day | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.ICU_StayLength | L4.ICU_Stay_7day | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.LengthOfStay | L4.LengthOfStay_3day | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.LengthOfStay | L4.LengthOfStay_7day | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.Mortality_Hospital | L4.ED_Inpatient_Mortality | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.Mortality_Hospital | L4.Inpatient_Mortality | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.Readmission_Hospital | L4.Readmission_30day | 40 |
+| ehr_bench | L1.text_only | L2.risk_prediction | L3.Readmission_Hospital | L4.Readmission_60day | 40 |
+| mm_bench | L1.multimodal | L2.cxr_vqa | L3.cxr_vqa_change | L4.cxr_change_comparison | 222 |
+| mm_bench | L1.multimodal | L2.cxr_vqa | L3.cxr_vqa_enumeration | L4.cxr_finding_enumeration | 220 |
+| mm_bench | L1.multimodal | L2.cxr_vqa | L3.cxr_vqa_presence | L4.cxr_finding_presence | 177 |
+| mm_bench | L1.multimodal | L2.phenotyping_mm | L3.Phenotype_ccs_mm | L4.phenotype_ccs | 120 |
+| mm_bench | L1.multimodal | L2.risk_prediction_mm | L3.Mortality_mm_full_stay | L4.inpatient_mortality_mm | 125 |
+| mm_bench | L1.multimodal | L2.risk_prediction_mm | L3.Mortality_mm_short_term | L4.mortality_24h | 125 |
+
+Subtotals: `ehr_bench` 45 leaves × 40 = 1,800 · `mm_bench` 989 across 6 leaves · grand total **2,789**.
+
+## Source data and provenance
+
+- **`ehr_bench`** is sub-sampled from the EHR-Bench benchmark of MIMIC-IV-derived clinical-decision and risk-prediction
+  tasks. The original 21,221 rows (13,500 decision_making + 7,721 risk_prediction) live under
+  `data/EHR-Bench/ehr_bench_decision_making.jsonl` and `data/EHR-Bench/ehr_bench_risk_prediction.jsonl` upstream;
+  the 1,800 rows here are the exact subset that was scored. The `qid` suffix equals the original `idx` in the upstream files.
+- **`mm_bench`** is sub-sampled from a combined multimodal test set (EHR-XQA-style CXR VQA +
+  CXR-conditioned mortality / phenotyping). The upstream file is
+  `data/EHR_multimodal_bench_tests/combined_test_set.jsonl` (2,703 rows), of which 989 are image-grounded; the remaining
+  1,704 text-only rows are excluded from this dataset.
+
+## Pairing with results
+
+The full evaluation outputs (per-row `f1 / precision / recall`, predictions, gold, `summary.json` aggregations at every
+taxonomy level) live in **`UCSC-VLAA/ClinSeek-Evaluation-Results`** under `hierarchical/{agentic,oneshot}/{bench}/{model}/`.
+The directory tree there mirrors the one here, replacing each `inputs.jsonl` with a `rows.jsonl` of scored predictions.
+
+## License & access
+
+Private. Inherits the access constraints of MIMIC-IV (PhysioNet credentialed) and MIMIC-CXR-JPG (PhysioNet credentialed)
+for the underlying patient data. Use only in accordance with those data-use agreements.

ehr_bench/all_inputs.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3832d1e127e88189f384278ad9765cb84c400c7335990c8099f898e8759dd7e9
+size 55786372