Datasets:

aditijc
/

dhf-mlm-baseline-eval-v1

code_id string	token string	n string	top1 string	top5 string	top1_acc string	top5_acc string
8	Glascow coma scale total	20320	13784	20320	0.6783464566929134	1.0
11	Heart Rate	12106	356	11171	0.029406905666611596	0.9227655707913431
15	Respiratory rate	11925	0	5074	0.0	0.4254926624737945
16	Systolic blood pressure	11925	1658	10085	0.1390356394129979	0.8457023060796646
4	Diastolic blood pressure	11868	1220	7063	0.10279743849005729	0.5951297607010448
14	Oxygen saturation	11832	0	4583	0.0	0.3873394185260311
13	Mean blood pressure	11781	1866	9853	0.15839062897886427	0.8363466598760716
17	Temperature	3342	0	3	0.0	0.0008976660682226212
6	Glascow coma scale eye opening	3132	217	986	0.06928480204342273	0.3148148148148148
9	Glascow coma scale verbal response	3064	7	38	0.0022845953002610967	0.012402088772845953
7	Glascow coma scale motor response	3046	52	418	0.017071569271175313	0.1372291529875246
10	Glucose	2654	0	8	0.0	0.003014318010550113
3	Capillary refill rate	1239	0	0	0.0	0.0
5	Fraction inspired oxygen	1003	0	0	0.0	0.0
19	pH	585	3	3	0.005128205128205128	0.005128205128205128
18	Weight	183	0	0	0.0	0.0
12	Height	1	0	0	0.0	0.0

dhf-mlm-baseline-eval-v1

Mar 30 retroactive eval of mlm_baseline.pt (pre-RACA, broken value_stats, no interpretability surface). Top-1 17.4% < most-frequent baseline 18.5%. Ablating value+dt improves to 18.2%. Fixed in v2.

Dataset Info

Rows: 17
Columns: 7

Columns

Column	Type	Description
code_id	Value('string')	Integer vocab ID for masked code
token	Value('string')	Raw vocab string (Glascow typo preserved)
n	Value('string')	Number of masked tokens of this code
top1	Value('string')	Top-1 correct count
top5	Value('string')	Top-5 correct count
top1_acc	Value('string')	top1 / n
top5_acc	Value('string')	top5 / n

Generation Parameters

{
  "script_name": "scripts/eval_mlm.py",
  "model": "mlm_baseline.pt",
  "description": "Mar 30 retroactive eval of mlm_baseline.pt (pre-RACA, broken value_stats, no interpretability surface). Top-1 17.4% < most-frequent baseline 18.5%. Ablating value+dt improves to 18.2%. Fixed in v2.",
  "experiment_name": "disentangled-health-futures",
  "cluster": "torch",
  "job_id": "torch:5132747",
  "artifact_status": "final",
  "canary": false,
  "split": "val",
  "batches": 25,
  "batch_size": 128,
  "ablation_summary": [
    {
      "mode": "none",
      "masked": "110006",
      "loss_ce": "2.3627546807105975",
      "top1": "0.1741995891133211",
      "top5": "0.6327382142792212",
      "mean_p_true": "0.11170084414962309"
    },
    {
      "mode": "no_value",
      "masked": "109920",
      "loss_ce": "2.3588081548654705",
      "top1": "0.1764919941775837",
      "top5": "0.6351710334788937",
      "mean_p_true": "0.11209837762202411"
    },
    {
      "mode": "no_dt",
      "masked": "110325",
      "loss_ce": "2.349969833446635",
      "top1": "0.1806571493315205",
      "top5": "0.6546023113528212",
      "mean_p_true": "0.1128443226489031"
    },
    {
      "mode": "no_diag",
      "masked": "110462",
      "loss_ce": "2.3657081486450093",
      "top1": "0.17444913182813998",
      "top5": "0.6310948561496261",
      "mean_p_true": "0.11073063418399469"
    },
    {
      "mode": "no_value_dt",
      "masked": "109544",
      "loss_ce": "2.343961432235563",
      "top1": "0.18211860074490616",
      "top5": "0.6579365369166728",
      "mean_p_true": "0.11332985960104358"
    }
  ],
  "ablation_metrics": {
    "none": {
      "masked": 110006,
      "loss_ce": 2.3627546807105975,
      "top1": 0.1741995891133211,
      "top5": 0.6327382142792212,
      "mean_p_true": 0.11170084414962309
    },
    "no_value": {
      "masked": 109920,
      "loss_ce": 2.3588081548654705,
      "top1": 0.1764919941775837,
      "top5": 0.6351710334788937,
      "mean_p_true": 0.11209837762202411
    },
    "no_dt": {
      "masked": 110325,
      "loss_ce": 2.349969833446635,
      "top1": 0.1806571493315205,
      "top5": 0.6546023113528212,
      "mean_p_true": 0.1128443226489031
    },
    "no_diag": {
      "masked": 110462,
      "loss_ce": 2.3657081486450093,
      "top1": 0.17444913182813998,
      "top5": 0.6310948561496261,
      "mean_p_true": 0.11073063418399469
    },
    "no_value_dt": {
      "masked": 109544,
      "loss_ce": 2.343961432235563,
      "top1": 0.18211860074490616,
      "top5": 0.6579365369166728,
      "mean_p_true": 0.11332985960104358
    }
  },
  "baseline_most_frequent": {
    "code_id": 8,
    "token": "Glascow coma scale total",
    "freq": 20320,
    "accuracy": 0.1847171972437867
  },
  "hyperparameters": {},
  "input_datasets": []
}

Usage

from datasets import load_dataset

dataset = load_dataset("aditijc/dhf-mlm-baseline-eval-v1", split="train")
print(f"Loaded {len(dataset)} rows")

Downloads last month: 50

Number of rows:

Total file size:

10.6 kB