Keep multimodal release source-only

README.md

@@ -20,328 +20,92 @@ configs:
 
 # ClinSeek-Bench
 
-The exact evaluation set used for the ClinSeek hierarchical benchmark
-(`UCSC-VLAA/ClinSeek-Evaluation-Results`, under `hierarchical/`). This
-repository ships **everything** needed to reproduce a run end to end:
+ClinSeek-Bench provides the evaluation manifests used by ClinSeekAgent. The
+multimodal split is released as a source-only manifest because MIMIC-derived
+patient databases, chest X-ray JPG files, radiology reports, and rendered EHR
+contexts are protected assets and must be rebuilt locally under each user's own
+credentialed access and data-use agreements.
 
-- the original benchmark inputs (questions, gold labels, patient pointers);
-- per-patient SQLite databases for every `subject_id` referenced;
-- the linked chest X-ray images and radiology reports for the multimodal
-  split;
-- table-description metadata used by the EHR MCP server.
+| Split | # data points | Modality | Released here |
+| --- | ---: | --- | --- |
+| `ehr_bench` | 1,800 | text-only EHR | evaluation manifest |
+| `mm_bench` | 989 | EHR + chest X-ray | source manifest only |
 
-| Split        | # data points | Modality                       | Total assets        |
-|--------------|--------------:|--------------------------------|---------------------|
-| `ehr_bench`  | **1,800**     | text-only                      | 1,563 patient DBs   |
-| `mm_bench`   | **989**       | text + chest X-ray (image)     | 560 patient DBs · 827 CXR JPGs · 356 radiology reports |
-| **Total**    | **2,789**     |                                |                     |
+`mm_bench` is restricted to image-grounded multimodal examples. It contains 497
+EHRXQA-derived CXR question-answering rows and 492 MedMod-derived ICU/CXR
+prediction rows.
 
-`mm_bench` here is restricted to **image-grounded** questions only — text-only EHR-query questions from the upstream
-multimodal test set are excluded because they are not part of the multimodal evaluation track.
+## Files
 
-## Two evaluation modes, one question set
-
-Each data point is evaluated under **two** modes that share an identical question set:
-
-| Mode      | Driver                          | Patient EHR delivered as       | Notes                                                                 |
-|-----------|---------------------------------|--------------------------------|-----------------------------------------------------------------------|
-| `agentic` | `deploy_agent.py` / `deploy_agent_mm.py` | **fetched at runtime** via an EHR MCP server (`ehr.load_ehr` + SQL/lookup tools) | Multi-turn tool use; agent decides what to look up. |
-| `oneshot` | `deploy_reasoning_model.py` / `deploy_reasoning_model_mm.py` | **inlined** in the prompt (pre-rendered EHR slice) | Single LLM call; everything the model needs is in the prompt. |
-
-The same `qid` produces one row in
-`UCSC-VLAA/ClinSeek-Evaluation-Results/hierarchical/agentic/<bench>/<model>/.../rows.jsonl`
-**and** one row in
-`UCSC-VLAA/ClinSeek-Evaluation-Results/hierarchical/oneshot/<bench>/<model>/.../rows.jsonl`,
-so the two scores are directly comparable.
-
-## How each mode reads the inputs
-
-### `ehr_bench`
-
-Both modes consume the same file: **`inputs/ehr_bench.json`** — a JSON list of 1,800 records. Each record carries fields
-for both modes:
-
-| Mode      | Fields the driver actually reads | Prompt the model sees |
-|-----------|----------------------------------|-----------------------|
-| `agentic` | `instruction`, `prediction_time`, `subject_id`, `candidates` | A small instruction wrapper that names the patient and lists candidate answers. **No patient EHR is inlined.** The agent must call `ehr.load_ehr(subject_id, prediction_time)` against the MCP server (which opens `data/ehr_bench/database/patient_<subject_id>.db`) and then issue further `ehr.*` tool calls. |
-| `oneshot` | `question` | A fully self-contained prompt (instruction + entire patient timeline rendered as Markdown) plus `<candidate_answers>` block. **No external lookup needed.** |
-
-Why two prompt forms? `question` is a long pre-rendered timeline (avg ≈10 KB, max ≈90 KB) — appropriate for one-shot
-inference but redundant in agentic mode where the agent should retrieve only what it needs.
-
-```python
-import json
-
-# Load the shared input file.
-rows = json.load(open("inputs/ehr_bench.json"))
-row  = next(r for r in rows if r["qid"] == "ehr_bench_decision_making_0")
-
-# --- agentic input (build the agent prompt + open the patient DB) -----------
-agentic_prompt_template = """The current time is {prediction_time}.
-The patient subject_id is {subject_id}.
-{instruction}
-<candidate_answers>
-{candidates}
-</candidate_answers>"""
-agentic_prompt = agentic_prompt_template.format(
-    prediction_time=row["prediction_time"],
-    subject_id=row["subject_id"],
-    instruction=row["instruction"],
-    candidates=json.dumps(row["candidates"]),
-)
-patient_db = f"data/ehr_bench/database/patient_{row['subject_id']}.db"   # the agent's MCP server opens this
-
-# --- oneshot input (single LLM call, prompt is fully self-contained) --------
-oneshot_prompt = row["question"]   # 1 kB → 90 kB, depending on the patient
-```
-
-The exact agentic prompt template is `EHR_Bench_Prompt` in
-[`openresearcher_ehr/data_utils.py:135`](https://github.com/Letian2003/DeepMed/blob/deepresearch/openresearcher_ehr/data_utils.py)
-(function `generate_ehr_bench_prompt`). The exact oneshot wrapper is
-[`build_user_content` in `deploy_reasoning_model.py:70`](https://github.com/Letian2003/DeepMed/blob/deepresearch/openresearcher_ehr/deploy_reasoning_model.py)
-(it appends a short `<answer>...</answer>` extraction trailer to `row["question"]`).
-
-### `mm_bench`
-
-Both modes consume the same file: **`inputs/mm_bench.jsonl`** — a JSONL of 989 records. Each record carries fields for
-both modes:
-
-| Mode      | Fields the driver actually reads | Prompt the model sees |
-|-----------|----------------------------------|-----------------------|
-| `agentic` | `question`, `image_paths`, `report_paths`, `subject_id`, `source_benchmark` | The bare task instruction (`question`, ≈540 chars). Patient EHR is fetched at runtime via EHR MCP (`ehr.load_ehr(subject_id, prediction_time)` against `data/mm_bench/<source>/database/patient_<subject_id>.db`); CXR images are attached to the user turn as image blocks; radiology reports may be inlined via tool calls. |
-| `oneshot` | `input_text` (preferred) or `question` (fallback), `image_paths` | A pre-rendered prompt (`input_text`, avg ≈24 KB, max ≈64 KB) that already contains `<patient_info>` + `<cxr_info>` + sliced `<ehr_context>` (top-N rows from each table) + `<image_inputs>` listing. CXR images are inlined into the user turn as base64 image blocks. **No DB lookup needed.** |
-
-```python
-import json
-
-# Load the shared input file.
-mm = [json.loads(l) for l in open("inputs/mm_bench.jsonl")]
-row = next(r for r in mm if r["qid"] == "medmod_radiology_test_56445431")
-
-# Pick the right bench root for assets and the per-patient DB.
-bench_root = f"data/mm_bench/{ 'ehrxqa' if row['source_benchmark']=='ehrxqa' else 'medmod' }"
-patient_db = f"{bench_root}/database/patient_{row['subject_id']}.db"
-
-# Resolve image_paths against bench_root (paths keep an upstream prefix that
-# we strip — see openresearcher_ehr/deploy_reasoning_model_mm.py:78).
-def resolve(p):
-    first, sep, rest = p.partition("/")
-    if sep and (first.endswith("OriginalLinked_v1") or first.endswith("AgentBench_v3")):
-        p = rest
-    return f"{bench_root}/{p}"
-
-image_files = [resolve(p) for p in (row.get("image_paths") or [])]
-
-# --- agentic input ----------------------------------------------------------
-agentic_prompt = row["question"]                            # ≈ 540 chars
-agentic_assets = {"images": image_files, "patient_db": patient_db}
-
-# --- oneshot input (preferred form) -----------------------------------------
-oneshot_prompt = row.get("input_text") or row["question"]   # ≈ 24 kB
-oneshot_assets = {"images": image_files}                    # no DB needed
-```
-
-`row["input_text"]` is the model-ready oneshot rendering; if it is missing or empty, fall back to `row["question"]`.
-The actual logic is `build_text_payload` in
-[`deploy_reasoning_model_mm.py:150`](https://github.com/Letian2003/DeepMed/blob/deepresearch/openresearcher_ehr/deploy_reasoning_model_mm.py).
-
-## Directory layout
-
-```
+```text
 ClinSeek-Bench/
-├── README.md                                  # this file
 ├── inputs/
-│   ├── ehr_bench.json                         # 1,800 records (text-only EHR-Bench)
-│   └── mm_bench.jsonl                         # 989 records (multimodal MM-Bench, image-grounded)
-└── data/
-    ├── ehr_bench/
-    │   ├── database/
-    │   │   ├── candidate_table.db             # vocabulary / candidate-answer table
-    │   │   └── patient_<subject_id>.db        # 1,563 per-patient SQLite DBs (≈2.0 GB)
-    │   └── table_description/                 # JSON metadata used by the EHR MCP server
-    │       ├── link_information.json
-    │       ├── shorten_description.json
-    │       └── table_description.json
-    └── mm_bench/
-        ├── ehrxqa/                            # rows where source_benchmark == "ehrxqa" (497 rows)
-        │   ├── database/                      # 165 per-patient SQLite DBs
-        │   ├── mimic-cxr/2.0.0/files/...      # linked CXR JPGs + report .txt
-        │   └── table_description/
-        └── medmod/                            # rows where source_benchmark == "medmod" (492 rows)
-            ├── database/                      # 395 per-patient SQLite DBs
-            ├── mimic-cxr/2.0.0/files/...      # linked CXR JPGs
-            └── table_description/
+│   ├── ehr_bench.json
+│   └── mm_bench.jsonl
+└── rebuild/
+    └── text_bench/
+        └── ClinSeek-Bench_text.json
 ```
 
-## Running an evaluation against this repo
+The multimodal manifest `inputs/mm_bench.jsonl` stores released `qid`s,
+questions, normalized labels, patient identifiers, prediction-time cutoffs, and
+relative CXR/report references. It intentionally does not store `input_text`,
+patient SQLite databases, CXR JPG files, or report text.
 
-After cloning, point each driver at the matching `--data` and `--bench-root` (or `--ehr_path`) inside this folder.
+## Reconstructing Multimodal Assets
 
-### EHR-Bench
+Multimodal reconstruction scripts are maintained in the ClinSeekAgent GitHub
+repository, matching the text benchmark data-preparation workflow:
 
-```bash
-# 0. Start the EHR MCP server on the ehr_bench root (agentic mode only)
-python src/run_mcp_server.py \
-    --mode http --host 127.0.0.1 --port 5103 \
-    --data_path /path/to/ClinSeek-Bench/data/ehr_bench
-# This server exposes ehr.load_ehr / ehr.run_sql_query / ehr.* tools backed
-# by the patient_*.db files in data/ehr_bench/database/.
-
-# 1a. Agentic — multi-turn tool use; reads instruction/subject_id/candidates,
-#     fetches EHR via the MCP server above.
-python openresearcher_ehr/deploy_agent.py \
-    --data_path /path/to/ClinSeek-Bench/inputs/ehr_bench.json \
-    --output_dir ./out_agentic_ehr \
-    --mcp_url http://127.0.0.1:5103/mcp \
-    --model <model-name>
+- documentation: https://github.com/UCSC-VLAA/ClinSeekAgent/blob/main/docs/ClinSeek-Bench_multimodal_data_prepare.md
+- scripts: https://github.com/UCSC-VLAA/ClinSeekAgent/tree/main/scripts/data_build
 
-# 1b. One-shot — single LLM call; reads `question` (which already inlines the
-#     patient timeline). No MCP server needed.
-python openresearcher_ehr/deploy_reasoning_model.py \
-    --data /path/to/ClinSeek-Bench/inputs/ehr_bench.json \
-    --output-dir ./out_oneshot_ehr \
-    --model <model-name>
-```
-
-### MM-Bench
+At a high level:
 
 ```bash
-# 0. Start the MM EHR MCP server on each multimodal sub-root (agentic only).
-#    The driver accepts multiple --bench-root paths so a single server isn't
-#    required, but the simplest setup is one MCP server per source split.
-python src/run_mcp_server.py --port 5113 --data_path /path/to/ClinSeek-Bench/data/mm_bench/ehrxqa &
-python src/run_mcp_server.py --port 5114 --data_path /path/to/ClinSeek-Bench/data/mm_bench/medmod &
-
-# 1a. Agentic — reads `question` + image_paths + subject_id; image MCP +
-#     EHR MCP servers handle assets and DB.
-python openresearcher_ehr/deploy_agent_mm.py \
-    --data /path/to/ClinSeek-Bench/inputs/mm_bench.jsonl \
-    --output_dir ./out_agentic_mm \
-    --bench_root /path/to/ClinSeek-Bench/data/mm_bench/ehrxqa \
-                 /path/to/ClinSeek-Bench/data/mm_bench/medmod \
-    --image_mcp_url http://127.0.0.1:5203/mcp \
-    --enable_ehr --ehrxqa_mcp_url http://127.0.0.1:5113/mcp \
-                 --medmod_mcp_url http://127.0.0.1:5114/mcp \
-    --model <model-name>
-
-# 1b. One-shot — reads `input_text` (already has EHR inlined) + image_paths.
-#     No MCP server needed; just resolve the images.
-python openresearcher_ehr/deploy_reasoning_model_mm.py \
-    --data /path/to/ClinSeek-Bench/inputs/mm_bench.jsonl \
-    --output-dir ./out_oneshot_mm \
-    --bench-root /path/to/ClinSeek-Bench/data/mm_bench/ehrxqa \
-                 /path/to/ClinSeek-Bench/data/mm_bench/medmod \
-    --model <model-name>
+git clone https://huggingface.co/datasets/UCSC-VLAA/ClinSeek-Bench
+git clone https://github.com/UCSC-VLAA/ClinSeekAgent.git
 ```
 
-The drivers above live in this repo:
-- <https://github.com/Letian2003/DeepMed/blob/deepresearch/openresearcher_ehr/deploy_agent.py>
-- <https://github.com/Letian2003/DeepMed/blob/deepresearch/openresearcher_ehr/deploy_reasoning_model.py>
-- <https://github.com/Letian2003/DeepMed/blob/deepresearch/openresearcher_ehr/deploy_agent_mm.py>
-- <https://github.com/Letian2003/DeepMed/blob/deepresearch/openresearcher_ehr/deploy_reasoning_model_mm.py>
-
-## Schema reference
-
-### `inputs/ehr_bench.json`
+Then run the scripts under `ClinSeekAgent/scripts/data_build/` with local paths
+to the official source datasets.
 
-```jsonc
-[
-  {
-    "qid": "ehr_bench_risk_prediction_329",          // stable id; matches results rows.jsonl
-    "subject_id": 12345678,                           // → patient DB lookup key
-    "hadm_id": 23456789,
-    "prediction_time": "2183-08-13 01:15:18",         // cutoff time the patient EHR is sliced at
-    "latest_event_time": "2183-08-13 01:15:17",
-    "task": "ED_Critical_Outcomes",                   // L4-style task tag
-    "task_type": "decision_making" | "risk_prediction",
-    "label": "yes" | ["..."],                         // gold answer
-    "question": "<task_instruction>...full timeline...</question>",   // ONESHOT prompt
-    "instruction": "Given the sequence of events ...",                // AGENTIC prompt
-    "input": "## Patient Demographics ...",                           // structured EHR (also fed into `question`)
-    "output": "OBSERVATION ADMIT",                                    // gold (string form)
-    "candidates": ["...", "..."],                                     // discrete answer options
-    "task_info": { "...": "..." },                                    // metric / target / split info
-    "source_idx": 0,
-    "source_file": "ehr_bench_risk_prediction.jsonl"
-  }
-]
-```
-
-### `inputs/mm_bench.jsonl`
+Required multimodal source data:
 
-```jsonc
-{
-  "qid": "medmod_radiology_test_56445431",          // stable id
-  "subject_id": 17795062,                            // → patient DB lookup key
-  "hadm_id": 22792685,
-  "stay_id": 39825171,
-  "prediction_time": "2177-07-25 22:04:23",
-  "task": "medmod_radiology",                        // task tag
-  "answer_type": "list" | "label",
-  "modalities": ["cxr_image", "cxr_table"],          // which assets are referenced
-  "ground_truth": [{"name": "Cardiomegaly"}, ...],   // gold answer
-  "question": "<task_instruction>...question only, no EHR...</question>",  // AGENTIC prompt
-  "input_text": "<task_instruction>...<ehr_context>...inlined CSV slices...<image_inputs>",  // ONESHOT prompt
-  "image_paths": ["MedModOriginalLinked_v1/mimic-cxr/2.0.0/files/p17/p17795062/s56445431/<dicom>.jpg"],
-  "report_paths": [],                                 // optional radiology report .txt
-  "source_benchmark": "medmod" | "ehrxqa",            // → which bench root to resolve assets against
-  "source_split": "test",
-  "source_index": 0
-}
-```
+| Dataset | Source |
+| --- | --- |
+| MIMIC-IV | https://physionet.org/content/mimiciv/ |
+| MIMIC-CXR | https://physionet.org/content/mimic-cxr/ |
+| MIMIC-CXR-JPG | https://physionet.org/content/mimic-cxr-jpg/ |
+| MIMIC-IV-Note | https://physionet.org/content/mimic-iv-note/ |
+| EHRXQA | https://physionet.org/content/ehrxqa/ |
+| MedMod code | https://github.com/nyuad-cai/MedMod |
 
-> **Path note.** `image_paths` and `report_paths` keep an upstream layout prefix (`MedModOriginalLinked_v1/...` or
-> `EHRXQAOriginalLinked_v1/...`). The drivers strip the first segment and re-resolve under each bench root, so the
-> physical files in this repo live at `data/mm_bench/<source>/mimic-cxr/2.0.0/...`. If you write your own loader,
-> mirror that strip-prefix-then-join logic.
+The final local reconstruction materializes `ClinSeek-MM-Bench/inputs/mm_bench.jsonl`
+with rendered `input_text`, plus `data/mm_bench/ehrxqa/` and
+`data/mm_bench/medmod/` containing patient SQLite databases, linked CXR JPGs,
+radiology reports when available, and table-description metadata.
 
-## Pairing with the result repository
+## Validation
 
-Per-row scored predictions live in **`UCSC-VLAA/ClinSeek-Evaluation-Results`** under
-`hierarchical/{agentic,oneshot}/{bench}/{model}/L1.<modality>/L2.<family>/L3.<sub>/L4.<task>/rows.jsonl`. To pull both
-modes for the same data point under one model:
+After cloning ClinSeekAgent, validate the released multimodal source manifest:
 
-```python
-import json, os
-
-INPUTS = "inputs/ehr_bench.json"     # or inputs/mm_bench.jsonl (read line-by-line)
-RESULTS_REPO = "/path/to/ClinSeek-Evaluation-Results"
-
-with open(INPUTS) as f:
-    rows = json.load(f)
-row = rows[0]                                          # any data point
-qid = row["qid"]
-
-def load_score(model, mode, bench):
-    base = f"{RESULTS_REPO}/hierarchical/{mode}/{bench}/{model}"
-    for dirpath, _, files in os.walk(base):
-        if "rows.jsonl" in files:
-            with open(os.path.join(dirpath, "rows.jsonl")) as f:
-                for line in f:
-                    r = json.loads(line)
-                    if r["qid"] == qid:
-                        return r
-    return None
-
-agentic = load_score("claude_opus_4_6", "agentic", "ehr_bench")
-oneshot = load_score("claude_opus_4_6", "oneshot", "ehr_bench")
-print("agentic F1:", agentic["f1"], "| oneshot F1:", oneshot["f1"])
+```bash
+python ClinSeekAgent/scripts/data_build/validate_multimodal_release.py \
+  --bench-root ClinSeek-Bench \
+  --manifest-only
 ```
 
-## Provenance
+Expected source-manifest counts:
 
-- **`ehr_bench`** is the 40-rows-per-task balanced subset of EHR-Bench. The full upstream files are
-  `data/EHR-Bench/ehr_bench_decision_making.jsonl` (13,500 rows) +
-  `data/EHR-Bench/ehr_bench_risk_prediction.jsonl` (7,721 rows) →
-  `ehr_bench_sampled_40_per_task.json` (1,800 rows = 27 decision_making leaves × 40 + 18 risk_prediction leaves × 40).
-- **`mm_bench`** is the image-grounded subset of the combined multimodal test set
-  (`combined_test_set.jsonl` / `model_ready_combined_test_set.jsonl`, 2,703 rows total). The 1,704 text-only
-  EHR-query rows are excluded; the 989 rows kept here are the rows scored under `mm_bench` in the result repository.
-- **EHR-Bench DBs** are sourced from the upstream EHR-Bench release (MIMIC-IV-derived per-patient snapshots).
-- **MM-Bench DBs / images / reports** are sourced from `EHRXQAAgentBench_v3` (EHR-XQA test rows) and
-  `MedModAgentBench_v3` (MedMod test rows) — both are MIMIC-IV / MIMIC-CXR-JPG derivatives.
+- 989 total multimodal rows.
+- 497 EHRXQA-derived rows.
+- 492 MedMod-derived rows.
+- 350 unique EHRXQA linked CXR JPG paths.
+- 477 unique MedMod linked CXR JPG paths.
+- 356 unique EHRXQA linked CXR report paths.
 
-## License & access
+## Citation
 
-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.
+Please cite the ClinSeek paper and the upstream source benchmarks used for any
+reconstructed multimodal assets.

data/mm_bench/ehrxqa/database/patient_10020740.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:e40190e2121733dce90354fc65b7e62b42f7d328fb2496b70260ff971df2765d
-size 110592

data/mm_bench/ehrxqa/database/patient_10117273.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:96e6711b3f43b9b74b3e4f474ad57d91bfa3f105ec9efb56f617fe71a087bac4
-size 294912

data/mm_bench/ehrxqa/database/patient_10156395.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:d8fefae661196a737541d75336f3deadd0be361e180a4c7ac983fd9740554ac9
-size 131072

data/mm_bench/ehrxqa/database/patient_10163947.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:be20ae3b09a0d2e18affa217621c10f04dd1519d0fa5d1b58ecf39205566f490
-size 139264

data/mm_bench/ehrxqa/database/patient_10176458.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:be59ade904bcb514641069ac77423a2fc60015a28d73c209207f47dd0a4584eb
-size 389120

data/mm_bench/ehrxqa/database/patient_10304606.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:8805d134a63d2ecf23db853ee9c73385e2eef72bf2eded78dc7a8b82664d99d4
-size 118784

data/mm_bench/ehrxqa/database/patient_10405915.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:b235800adbab14163e8d95f5493f066b656e2f6a2b4a8d881aa6ecfa049407b4
-size 114688

data/mm_bench/ehrxqa/database/patient_10410672.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:53a20b617bb5d5b3db72f7311e5d031ab13c5b0ab59879cdceac0052c8f0641c
-size 131072

data/mm_bench/ehrxqa/database/patient_10501557.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:33dad98f6f5d8c9bcb31be565d7c1750417e7f82ff63a79bb7e166c7ab9ea086
-size 118784

data/mm_bench/ehrxqa/database/patient_10518993.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:abc9dc26fa13e24aa18911874af0e5f23dc20158b246569557da737ac0bf5a4f
-size 417792

data/mm_bench/ehrxqa/database/patient_10680329.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:c3b6a02b4433cce33ef8275705c760e8f7579e5b6c8605d6816377b43209dd54
-size 126976

data/mm_bench/ehrxqa/database/patient_11053589.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:0f8287822293919376d1b504a47831e2f14d28eddd8d7161735b239500e31ee8
-size 233472

data/mm_bench/ehrxqa/database/patient_11111102.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:78553d92582c3bf9befa55b2f9794cf516ed07363dcb7900cb9cec05ed5ea473
-size 278528

data/mm_bench/ehrxqa/database/patient_11320106.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:ff78be6ec908772702ae4a5aef49b1a0f704fae558d79ce143b7fd422112ec64
-size 172032

data/mm_bench/ehrxqa/database/patient_11350319.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:61a6f2b87dd35771b6072c42442adf0c24da55257927b01c76f61a6264b56e9c
-size 286720

data/mm_bench/ehrxqa/database/patient_11422357.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:7a1dc603a6839db2d1794c1ae62f46a49383ccd8a92dd732c8e1a6af93f7b0de
-size 159744

data/mm_bench/ehrxqa/database/patient_11572460.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:c9e3922f49279a0a5ea7dfcb5f2cfc20df6652d46696c5ce5cde9f490835aea9
-size 110592

data/mm_bench/ehrxqa/database/patient_11632359.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:42c15c1673b400ad7f2e8e3e29f4dd30f705e3a6c68344b282a853f262a88b80
-size 110592

data/mm_bench/ehrxqa/database/patient_11636169.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:d20ea570399e6dc2211fd9b1896fc705e46abbb792cad3bdad63065e7ea768ed
-size 221184

data/mm_bench/ehrxqa/database/patient_11801290.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:48df3ea8d1f5420c2b36ee5ad10d80691949aae98fa77b92f54f9ed122c11dc4
-size 176128

data/mm_bench/ehrxqa/database/patient_11929103.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:bfa29c1228158db1ecdf36f006ed72f1bd941189bf996606ae9eed2ffc86077a
-size 163840

data/mm_bench/ehrxqa/database/patient_11950244.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:4aab78df9dc9eda9a77e5de7fdcc3d07ea9076e11d19fe0b78911a730a41ae9d
-size 159744

data/mm_bench/ehrxqa/database/patient_11966699.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:c29205befe682161c087119bdae30d911b1031bbf59b56b0dd27b70e10e2bef9
-size 192512

data/mm_bench/ehrxqa/database/patient_11982468.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:d29d48cec5615ed2bd1e8783131453a1de6fd88a087916869455150583f66e6c
-size 118784

data/mm_bench/ehrxqa/database/patient_12106566.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:d3f2749ee2086c093e7f3fb609ffa68a57c7a97e068f864a26a003513c390ed9
-size 110592

data/mm_bench/ehrxqa/database/patient_12215941.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:ce05444d57ecff94878f1abaecb80a7c0ecb3d834d7d8d04801a68c41bfe48c3
-size 356352

data/mm_bench/ehrxqa/database/patient_12250782.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:6f85ff4f7b9e2407bcf48860924a27aaa5ecae7ec39edc70196d33b024718024
-size 131072

data/mm_bench/ehrxqa/database/patient_12262788.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:9814dbaff706089131b5134cae640641766b6e8b933c66a46de93db719d21a82
-size 118784

data/mm_bench/ehrxqa/database/patient_12360014.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:a3c91aab006db904cf85c4fc921c20625b485df21a98f3b425208fd09dff32d2
-size 135168

data/mm_bench/ehrxqa/database/patient_12405140.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:4d43ab30f8473e8aacace0322085bf16fb1ebb658ae930b12917ee75804cf1f8
-size 126976

data/mm_bench/ehrxqa/database/patient_12429112.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:b9fead6b2d922f034e6f7524cf3f42a57fd53d428bf13a8b900bf237e7b1cc68
-size 155648

data/mm_bench/ehrxqa/database/patient_12604446.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:6650cc0b7d94c9c103d1e9d85111e9caeef5bff0af0db5910a66205ac5d3961a
-size 172032

data/mm_bench/ehrxqa/database/patient_12648153.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:320f5730e814070e19aa3cfeed00c2aebae999f773b789dc26035e93b01dedb5
-size 204800

data/mm_bench/ehrxqa/database/patient_12716528.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:b7d483a20097dc9d588f9f9eee3abddd0be19e5112480d4f8697b2416bda377f
-size 868352

data/mm_bench/ehrxqa/database/patient_12724975.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:77780a2c723ef6114dc6b504163ff6d2b4a8ac4120887c82feb31c55015c6548
-size 315392

data/mm_bench/ehrxqa/database/patient_12727273.db

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:bd99d5a40dc7cffe2ff1a390447f111589281f26a4eb39eaba5edae391944e9b
-size 139264