PED-X-Bench: FDA Pediatric Drug Extrapolation Dataset

This repository contains the dataset, benchmark tasks, and baseline models for ICLR 2026.

PED-X-Bench: A Dataset for Modeling FDA Pediatric Drug Extrapolation Decisions

🧾 Overview

PED-X-Bench is a benchmark for evaluating models on the task of predicting whether the U.S. FDA extrapolated adult drug data to children in labeling decisions. It includes:

• ✅ 778 structured FDA drug label entries (2007–2024)
• ✅ Extrapolation labels: Full, Partial, None, Unlabeled
• ✅ Summaries of pediatric efficacy and PK/safety evidence
• ✅ Annotated rationales and pediatric study characteristics
• ✅ Manually adjudicated subset of 135 entries

This creates the exact directory layout expected by train_bigbird.py.

Quick-start: reproduce the BigBird baseline

1. Create a clean environment

conda create -n pedx-bench python=3.10 -y conda activate pedx-bench pip install -r requirements.txt # transformers[torch], datasets, accelerate, evaluate, scikit-learn, sentencepiece

2. Train for four epochs (≈20 min on 1 × A100; CPU works but is slower)

python scripts/train_bigbird.py \
       --split_dir data/processed/splits \
       --txt_dir   data/raw/txt \
       --out_dir   checkpoints/bigbird_demo \
       --epochs    4

The script prints dev metrics every 100 steps and writes: checkpoints/bigbird_demo/ ├── config.json ├── pytorch_model.bin ├── tokenizer.json └── test_metrics.json

Evaluate the saved model

python scripts/eval_bigbird.py \
       --model_dir checkpoints/bigbird_demo \
       --split_csv data/processed/splits/test.csv \
       --txt_dir   data/raw/txt