HEARTS Data Samples

This repository hosts the fixed ("frozen") test cases used by the HEARTS benchmark. Each file is a Python pickle (.pkl) containing one test-case payload.

Quick Links

• Dataset (this repo): https://huggingface.co/datasets/yang-ai-lab/HEARTS
• Code (HEARTS framework): https://github.com/yang-ai-lab/HEARTS

Contents

• HEARTS Data Samples
  • Quick Links
  • Contents
  • Folder Layout
  • Download
  • Use with HEARTS
  • Inspect / Load a Sample
  • Notes

Folder Layout

All samples follow this pattern:

<dataset>/<task>/*.pkl

Example:

<root>/
  cgmacros/
    cgm_stat_calculation/
      0.pkl
      1.pkl
  vitaldb/
    some_task/
      0.pkl

Conventions:

• <dataset>: dataset identifier (e.g., cgmacros, vitaldb)
• <task>: task identifier within that dataset
• N.pkl: N-th fixed test case for that dataset/task

Download

Option A: Clone (recommended for binary files; may require Git LFS)

git lfs install
git clone https://huggingface.co/datasets/yang-ai-lab/HEARTS

Option B: Download programmatically (Python)

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="yang-ai-lab/HEARTS",
    repo_type="dataset",
)
print("Downloaded to:", local_dir)

Tip: Download only a subset via allow_patterns (reduces disk/network)

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="yang-ai-lab/HEARTS",
    repo_type="dataset",
    allow_patterns=[
        "cgmacros/cgm_stat_calculation/*.pkl",
    ],
)
print("Downloaded to:", local_dir)

Use with HEARTS

In the HEARTS codebase, the root directory you downloaded/cloned is the fix_test_cases_dir.

uv run run_exp_freeze.py \
  --fix-test-cases-dir /path/to/HEARTS \
  --dataset-name cgmacros \
  --task cgm_stat_calculation

Inspect / Load a Sample

Security note: Python pickles can execute code when loaded. Only unpickle files you trust.

import pickle
from pathlib import Path

root = Path("/path/to/HEARTS")
sample_path = root / "<dataset>" / "<task>" / "0.pkl"

with sample_path.open("rb") as f:
    obj = pickle.load(f)

print(type(obj))
print(obj)

List available datasets/tasks:

from pathlib import Path

root = Path("/path/to/HEARTS")

for dataset_dir in sorted([p for p in root.iterdir() if p.is_dir()]):
    task_dirs = sorted([p for p in dataset_dir.iterdir() if p.is_dir()])
    print(dataset_dir.name, "tasks:", [p.name for p in task_dirs])

Notes

• These files are intended to be immutable inputs for reproducible evaluation.
• Pickled objects can be code/version-dependent; if you see load/format issues, align your HEARTS code version with the data release.