Datasets:

omicsbank
/

eval-harness-demo

patient_id stringlengths 9 9	age int64 1 94	gender stringclasses 3 values	num_encounters int64 1 30	num_clinical_notes int64 1 104	has_imaging_study bool 2 classes	primary_imaging_modality stringclasses 5 values	has_lab_results bool 2 classes	record_span_years float64 0.2 3.9
EVAL-0001	89	Male	3	12	true	CT	true	3.1
EVAL-0002	59	Male	8	31	true	X-Ray	false	2.9
EVAL-0003	65	Male	4	16	false	null	true	3.4
EVAL-0004	85	Male	4	12	false	null	false	2.8
EVAL-0005	55	Female	3	12	true	MRI	true	3
EVAL-0006	73	Male	7	11	true	MRI	false	1.3
EVAL-0007	79	Male	1	1	true	MRI	true	0.8
EVAL-0008	22	Male	4	6	false	null	false	3.1
EVAL-0009	6	Female	1	2	false	null	true	0.7
EVAL-0010	29	Male	2	4	true	X-Ray	false	3.7
EVAL-0011	27	Male	17	35	true	X-Ray	true	2.4
EVAL-0012	83	Female	5	16	false	null	true	1.4
EVAL-0013	86	Male	2	3	true	X-Ray	true	3.8
EVAL-0014	1	Female	2	6	true	X-Ray	false	0.7
EVAL-0015	47	Female	9	17	true	MRI	false	2.1
EVAL-0016	78	Female	2	7	false	null	true	0.5
EVAL-0017	13	Male	5	7	true	CT	false	3.9
EVAL-0018	75	Female	3	7	false	null	false	2.3
EVAL-0019	12	Male	2	8	false	null	false	3.2
EVAL-0020	44	Female	2	6	true	X-Ray	true	1.2
EVAL-0021	77	Male	6	23	true	CT	true	3.2
EVAL-0022	29	Female	8	27	false	null	true	2.8
EVAL-0023	33	Male	1	3	true	MRI	true	2.6
EVAL-0024	27	Male	17	35	true	X-Ray	true	3.8
EVAL-0025	68	Male	3	9	true	MRI	false	1.8
EVAL-0026	24	Male	7	11	true	Ultrasound	true	1.7
EVAL-0027	94	Female	6	23	false	null	true	2.8
EVAL-0028	42	Male	1	4	true	X-Ray	false	3.4
EVAL-0029	45	Other/Unspecified	4	16	true	X-Ray	false	1.4
EVAL-0030	48	Female	6	9	false	null	true	2.7
EVAL-0031	55	Female	3	7	false	null	false	0.9
EVAL-0032	53	Female	4	9	true	MRI	true	2.3
EVAL-0033	48	Female	5	9	false	null	false	3.1
EVAL-0034	94	Male	3	10	false	null	true	0.4
EVAL-0035	76	Male	6	23	true	Other	false	2.9
EVAL-0036	75	Male	2	4	false	null	false	0.2
EVAL-0037	66	Male	2	8	true	X-Ray	false	3.8
EVAL-0038	59	Male	2	7	false	null	true	1.9
EVAL-0039	33	Female	1	1	false	null	false	1.7
EVAL-0040	93	Male	3	11	false	null	false	2.9
EVAL-0041	44	Female	17	70	false	null	true	2.1
EVAL-0042	21	Male	4	8	true	MRI	true	3
EVAL-0043	80	Female	1	3	true	MRI	true	0.4
EVAL-0044	16	Female	2	6	false	null	false	2.3
EVAL-0045	81	Male	16	53	false	null	false	2.3
EVAL-0046	58	Female	3	5	true	Ultrasound	true	3.7
EVAL-0047	11	Female	30	104	true	X-Ray	false	0.3
EVAL-0048	5	Female	2	6	false	null	true	1.9
EVAL-0049	42	Female	4	15	false	null	false	2.6
EVAL-0050	4	Male	8	31	false	null	false	2.2
EVAL-0051	14	Male	7	17	true	X-Ray	false	0.4
EVAL-0052	49	Female	4	14	true	Other	true	2.4
EVAL-0053	92	Female	2	8	false	null	true	1
EVAL-0054	44	Male	10	41	false	null	false	0.9
EVAL-0055	76	Female	4	7	false	null	true	3.5
EVAL-0056	87	Female	17	26	false	null	true	0.9
EVAL-0057	78	Female	3	10	false	null	false	1.9
EVAL-0058	60	Male	4	8	false	null	false	3
EVAL-0059	42	Male	4	12	false	null	false	2.9
EVAL-0060	49	Male	15	65	true	X-Ray	false	2.3

OmicsBank Eval Harness (Demo)

⚠️ This uses entirely synthetic data. No row corresponds to a real patient. Built to demonstrate the harness mechanics — not to represent real-world statistics or actual clinical patterns.

Two tools, both runnable in a couple minutes with just pandas installed:

Auto-grader — score your model's predictions on a small held-out classification task
Leakage checker — check whether your eval/training data overlaps with a reference corpus

Quick start

git clone https://huggingface.co/datasets/omicsbank/eval-harness-demo
cd eval-harness-demo
pip install -r requirements.txt

1. Auto-grader

# Run the baseline (majority-class) predictor to see the pipeline work end to end:
python predict_baseline.py
python auto_grader.py --predictions predictions.csv

# Then swap in your own model — write predictions to a CSV with columns
# patient_id, diagnosis_category (using data/eval_features.csv as input),
# and grade it the same way:
python auto_grader.py --predictions your_predictions.csv

The task: predict diagnosis_category (5 classes) from patient/encounter features in data/eval_features.csv. Ground truth lives in data/ground_truth.csv — the auto-grader merges on patient_id and reports accuracy, macro F1, and per-class precision/recall/F1.

2. Leakage checker

# Run against the bundled demo pair first, to see how it works:
python leakage_checker.py --id-col record_id --text-col note_text

# Then point it at your own files:
python leakage_checker.py --file-a your_eval_set.csv --file-b public_corpus.csv \
    --id-col <id_column> --text-col <text_column>

Runs two checks: exact ID overlap, and near-duplicate text similarity (catches paraphrased/lightly-edited duplicates an exact match would miss). Either flag being non-trivial is a sign your benchmark results may be inflated by contamination.

Why this exists

Most teams building on public clinical data haven't checked whether their eval set overlaps with what they (or anyone else) trained on. This harness is a small, honest demonstration of that check — and of what a genuinely graded, held-out benchmark task looks like end to end.

The data here is synthetic and small on purpose. OmicsBank's actual licensed dataset — longitudinal, multimodal EMR data (India), including a genuinely held-out eval set that has never been public — is available under a commercial data license.

Contact: sales@omicsbank.com — include your company, role, and intended use case.

Downloads last month: 19

Number of rows:

Total file size:

22.3 kB