MediSimplifier Dataset

Overview

Instruction-formatted dataset for medical discharge summary simplification, derived from Asclepius-Synthetic-Clinical-Notes with Claude-generated ground truth.

Source

• Original: starmpcc/Asclepius-Synthetic-Clinical-Notes (158K total)
• Used: 10,000 samples
• Ground Truth: Generated using Claude Opus 4.5 (claude-opus-4-5-20251101)

Splits

Statistics

• Length: 779-5,208 characters (mean: 1,933, median: 1,806)
• Target F-K Grade: ~7.2 (simplified ground truth)

Format

Arrow format with HuggingFace datasets library:

from datasets import load_from_disk

dataset = load_from_disk("data/")
print(dataset)
# DatasetDict({
#     train: Dataset({features: ['instruction', 'input', 'output'], num_rows: 7999}),
#     validation: Dataset({features: ['instruction', 'input', 'output'], num_rows: 999}),
#     test: Dataset({features: ['instruction', 'input', 'output'], num_rows: 1001})
# })

Fields

• instruction: System prompt for simplification task
• input: Original complex medical discharge summary
• output: Simplified version (Claude-generated ground truth)

Usage

from datasets import load_from_disk

dataset = load_from_disk("data/")

# Access a sample
sample = dataset["train"][0]
print(f"Input: {sample['input'][:200]}...")
print(f"Output: {sample['output'][:200]}...")

License

Dataset derived from Asclepius-Synthetic-Clinical-Notes under its original terms. Ground truth simplifications generated via Anthropic API.