language:

• en
  license: mit
  pretty_name: DAG Remediation Traces
  tags:
• phi
• de-identification
• clinical-nlp
• privacy
• dag
• remediation-planning
• graph
• planning
• healthcare
• hipaa
• multimodal
• synthetic
• budget-constrained
• topology
• dependency-injection
  size_categories:
• 1K<n<10K
  task_categories:
• text-classification
• other
  configs:
• config_name: default
  data_files:
  • split: train
    path: data/train.jsonl
  • split: test
    path: data/test.jsonl
• config_name: hard
  data_files:
  • split: test
    path: data/hard.jsonl

DAG Remediation Traces

Author: Venkata Krishna Azith Teja Ganti
Part of the ExposureGuard PHI Re-identification Risk Ecosystem

Input/output execution traces for budget-constrained PHI remediation planning over multimodal clinical records. Each record pairs a patient risk profile with a complete DAG planning trace: which actions were selected, which dependency injections fired, the topological execution order, and the final residual risk and cost.

Use this to train or benchmark alternative remediation planners, study cost/risk tradeoffs, or evaluate DAG construction under budget constraints across text, image, audio, and EHR modalities.

Splits

The hard config fixes input risk between 0.75 and 0.99 and caps budget at 0.60, creating a stress-test set where every record requires planning under tight constraints.

Plan quality

Numbers are shown for all records and for planned-only (records where the planner actually ran). The gap between the two reflects the 15.8% of default train records that were already below the risk threshold and received empty plans. The hard split has no below-threshold records so both columns are the same.

The injected deps rate nearly doubles in the hard split (33.6% to 67.9%) -- tighter budgets under high risk force more complex dependency chains.

These metrics are self-reported by the generator. They describe how the planner performs against its own scoring logic, not against externally validated PHI risk reduction.

Input risk distribution (default train)

Input risk distribution (hard)

Quick Start

from datasets import load_dataset

ds = load_dataset("vkatg/dag_remediation_traces")
train = ds["train"]
test  = ds["test"]

import json

with open("data/train.jsonl") as f:
    records = [json.loads(line) for line in f]

planned = [r for r in records if r["status"] == "plan_ready"]
print(len(planned), "planned records")
print("avg residual risk:", sum(r["estimated_residual_risk"] for r in planned) / len(planned))

Schema

Action Catalog (14 actions)

Dependency graph

Some actions require others to run first:

• retokenize_text requires mask_direct_id
• k_anon_table requires generalize_dob, suppress_geo, drop_rare_code
• cross_modal_unlink requires redact_image_face, strip_audio_voice, mask_direct_id
• federated_noise requires k_anon_table, cross_modal_unlink
• audit_log_purge requires federated_noise
• redact_image_tag requires redact_image_face
• anon_audio_content requires strip_audio_voice

When a selected action has unmet dependencies, those dependencies are injected automatically. 33.6% of planned records include at least one injected dependency.

Action frequency (train)

Sample Record

{
  "patient_id": "P-000001",
  "input": {
    "risk_score": 0.9051,
    "retok_prob": 0.2147,
    "active_modalities": ["ehr"],
    "budget": 0.86
  },
  "plan": [
    {"action": "generalize_dob", "priority": 8.1366, "cost": 0.03, "risk_delta": 0.12, "deps": []},
    {"action": "suppress_geo",   "priority": 5.0854, "cost": 0.04, "risk_delta": 0.10, "deps": []},
    {"action": "cross_modal_unlink", "priority": 3.1642, "cost": 0.18, "risk_delta": 0.28, "deps": ["redact_image_face", "strip_audio_voice", "mask_direct_id"]}
  ],
  "trace": {
    "selected_actions": ["generalize_dob", "suppress_geo", "cross_modal_unlink", "drop_rare_code", "audit_log_purge", "k_anon_table"],
    "injected_deps": ["redact_image_face", "strip_audio_voice", "mask_direct_id", "federated_noise"],
    "topo_order": ["generalize_dob", "suppress_geo", "redact_image_face", "strip_audio_voice", "mask_direct_id", "cross_modal_unlink", "drop_rare_code", "federated_noise", "audit_log_purge", "k_anon_table"],
    "score_breakdown": {"generalize_dob": 8.1366, "suppress_geo": 5.0854, "cross_modal_unlink": 3.1642}
  },
  "estimated_residual_risk": 0.0,
  "total_cost": 1.1,
  "status": "plan_ready"
}

Planner Logic

The scorer ranks actions by a return-on-investment signal:

score = (reduces * modality_weight + retok_bonus) / base_cost * log(1 + risk * 5)

modality_weight is sampled per record (0.8 to 1.5) for a random subset of active modalities, reflecting variable sensitivity across data types. retok_bonus (+0.15) fires for retokenize_text when retok_prob > 0.55.

Selection adds actions greedily by score until the budget is hit or the risk reduction target is met (target = max(0, risk - 0.20), minimum spend 0.10). Dependency injection runs after selection to ensure topological validity.

Reproduce

python generate_dag_traces.py

The default train/test splits are deterministic at seed 7. The hard split uses seed 42 with risk sampled from [0.75, 0.99] and budget from [0.30, 0.60].

Related Models

Related Datasets

• vkatg/streaming-phi-deidentification-benchmark
• vkatg/phi-audit-trace-benchmark

Citation

@misc{ganti2025exposureguard,
  title        = {ExposureGuard: Cross-Modal PHI Re-identification Risk Scoring with DCPG and Federated CRDT Distillation},
  author       = {Ganti, Venkata Krishna Azith Teja},
  year         = {2025},
  doi          = {10.5281/zenodo.18865882},
  howpublished = {\url{https://huggingface.co/vkatg}}
}

License

MIT. Fully synthetic data. Contains no real patient information.