README.md

LongHealth Benchmark - Preprocessed for Packed Attention

This dataset is a preprocessed version of the LongHealth benchmark, formatted for evaluation with packed document attention models.

Dataset Description

LongHealth is a medical question-answering benchmark that tests long-context understanding across multiple clinical documents. Each example consists of multiple medical documents (notes, lab reports, discharge summaries) and a multiple-choice question.

Format

The dataset is a JSON file with the following structure:

{
  "metadata": {
    "source": "benchmark_v5.json",
    "tokenizer": "google/t5gemma-2b-2b-prefixlm-it",
    "max_encoder_len": 55000,
    "add_distractors": true,
    "preprocessing_timestamp": "2024-12-02T...",
    "num_patients": X,
    "num_questions": Y,
    "stats": { ... }
  },
  "examples": {
    "patient_001_q00": {
      "answer_document_tokens": [[101, 2023, ...], [101, 5011, ...]],
      "distractor_document_tokens": [[101, 8392, ...], ...],
      "decoder_input_ids": [1, 2, 3, ...],
      "question_text": "What was the primary diagnosis?",
      "correct_answer": "A",
      "correct_text": "Pneumonia",
      "all_options": {
        "A": "Pneumonia",
        "B": "Bronchitis",
        "C": "...",
        "D": "...",
        "E": "..."
      },
      "patient_id": "patient_001",
      "question_idx": 0,
      "answer_doc_ids": ["text_0", "text_1"],
      "num_answer_docs": 2,
      "num_distractor_docs": 8,
      "total_context_length": 12453,
    }
  }
}

Field Descriptions

Example Fields

• answer_document_tokens: List of tokenized documents containing the answer (must be included)
• distractor_document_tokens: List of tokenized distractor documents (included if budget allows)
• decoder_input_ids: Tokenized decoder prompt with chat template applied
• question_text: Original question text
• correct_answer: Correct answer letter (A, B, C, D, or E)
• correct_text: Full text of the correct option
• all_options: Dictionary mapping letters to option texts
• patient_id: ID of the patient case
• question_idx: Index of question within patient case (0-19)
• answer_doc_ids: Original document IDs that contain the answer
• num_answer_docs: Number of answer documents included
• num_distractor_docs: Number of distractor documents included
• total_context_length: Total tokens in encoder context
• budget_exceeded: Whether answer documents exceeded max_encoder_len

Metadata Stats

The metadata includes comprehensive statistics:

• Document length distributions (min/max/mean/median/p95)
• Answer document totals per question
• Final context lengths after budget application
• Number and percentage of questions where budget was exceeded
• Average number of answer/distractor documents per question

Token Budget Logic

1. Answer documents are prioritized - they are always included first
2. If total answer document length > max_encoder_len:
   • Keep first N answer documents that fit within budget
   • No distractor documents are added
   • budget_exceeded = True
3. If answer documents fit:
   • All answer documents included
   • Distractor documents added greedily until budget is full (if add_distractors=True)
   • budget_exceeded = False

Decoder Prompt Format

The decoder prompt uses the model's chat template with add_generation_prompt=True:

Answer this multiple choice question based on the medical documents provided in the context.

Question: {question}
A: {option_a}
B: {option_b}
C: {option_c}
D: {option_d}
E: {option_e}

You must respond in this exact format:
'The correct answer is [LETTER]: [Full option text]'

Example: 'The correct answer is B: Acute bronchitis.'

Usage

For evaluation with packed attention models:

import json
import torch
from transformers import AutoTokenizer

# Load dataset
with open("longhealth_preprocessed.json") as f:
    data = json.load(f)

# Process example
example = data["examples"]["patient_001_q00"]

# Prepare for model
encoder_inputs = [
    torch.tensor(tokens, dtype=torch.long)
    for tokens in example["answer_document_tokens"]
] + [
    torch.tensor(tokens, dtype=torch.long)
    for tokens in example["distractor_document_tokens"]
]

decoder_input = torch.tensor(example["decoder_input_ids"], dtype=torch.long)

# Evaluate response
def is_correct(generated_text, correct_letter, correct_text):
    return (correct_letter in generated_text and 
            correct_text.lower() in generated_text.lower())

Citation

If you use this dataset, please cite the original LongHealth paper:

[LongHealth citation to be added]

License

Same as original LongHealth benchmark.