chrisvoncsefalvay/vaers-narrative-generation

Adverse Event Symptom Narrative Generation Dataset

Dataset description

A dataset of VAERS symptom narratives and diagnostic (laboratory) findings transformed into instruction-tuning format (Alpaca) to finetune conditional and unconditional generative models.

This dataset contains 2,649,322 training examples derived from VAERS (Vaccine Adverse Event Reporting System) data for the years 1990-2025.

The data has been transformed into instruction-tuning format (Alpaca) suitable for fine-tuning language models on clinical adverse event reporting.

Dataset summary

• Task: Clinical text generation from structured patient and vaccine information
• Language: English
• Output field: SYMPTOM_TEXT
• Patient history included: No
• Format: Alpaca instruction-tuning format
• Years covered: 1990-2025
• Total examples: 2,649,322

Key features

• Automatic deduplication: One training example per VAERS_ID (consolidates multiple vaccines per patient)
• Manufacturer exclusion: Vaccine manufacturers excluded to avoid brand bias
• List preservation: Vaccines and symptoms preserved as lists for multiple-value fields
• Quality filtering: Records with empty output fields excluded

Data structure

Each example contains three fields following the Alpaca format:

• instruction: Task description for the model
• input: Structured patient and vaccine information including:
  • Age (years)
  • Sex
  • Vaccine(s) administered (multiple vaccines joined with commas)
  • Symptoms (multiple symptoms joined with commas, MedDRA versions filtered)
  • Patient history (optional, if include_history=True)
• output: Clinical narrative text from VAERS reports

Data fields

Field	Type	Description
instruction	string	Task instruction for the model
input	string	Structured patient/vaccine information (newline-separated)
output	string	Clinical narrative (SYMPTOM_TEXT)

Internal data representation

While the published dataset uses Alpaca format (string fields), the internal processing preserves structured data:

• vaccines_list: List of vaccine types (e.g., ["COVID19", "FLU"])
• symptoms_list: List of symptoms (e.g., ["Headache", "Fatigue", "Myalgia"])
• manufacturers_list: List of manufacturers (excluded from training data)
• dose_series_list: List of dose numbers

Example Record

{
  "instruction": "Generate a clinical symptom description based on the patient and vaccine information provided",
  "input": "Age: 0.2\nSex: F\nVaccine: DTP\nSymptoms: Agitation",
  "output": "Loud intense cry with screaming for 1 1/2 hrs.  Seen next day, child  normal."
}

Source data

This dataset is derived from the CDC VAERS (Vaccine Adverse Event Reporting System) public data, available at: https://vaers.hhs.gov/data.html

VAERS is a national early warning system to detect possible safety problems in U.S.-licensed vaccines. The system is co-managed by the CDC and FDA.

Data processing

The raw VAERS data consists of three CSV files per year:

• VAERSDATA: Patient demographics and clinical narratives
• VAERSVAX: Vaccine administration details
• VAERSSYMPTOMS: Coded symptoms using MedDRA terminology

Processing steps:

1. Loading: CSV files loaded with automatic encoding detection (supports UTF-8, Latin-1, Windows-1252)
2. Joining: Tables joined by VAERS_ID
3. Deduplication: Multiple vaccines per patient consolidated into single training example
4. Filtering: MedDRA version numbers removed, empty outputs excluded, manufacturers excluded
5. Transformation: Converted to Alpaca instruction-tuning format

Intended use

Primary use cases

• Fine-tuning language models for clinical adverse event reporting
• Training models to generate clinical narratives from structured patient data
• Research in medical natural language generation
• Development of clinical documentation assistance tools

Out-of-scope use

• Clinical decision-making without expert oversight
• Automated diagnosis or treatment recommendations
• Replacement for professional medical judgment

Limitations and biases

Data limitations

• Reporting bias: VAERS is a passive surveillance system; not all adverse events are reported
• Causality: VAERS reports do not establish causal relationships between vaccines and adverse events
• Completeness: Not all fields are complete in every report
• Temporal coverage: Dataset covers years 1990-2025

Potential biases

• Reporting patterns: Healthcare providers and vaccine manufacturers are required to report certain events, while patient reporting is voluntary
• Media influence: Increased reporting following media coverage of vaccine-related events
• Temporal bias: Reporting practices and data quality have evolved over time

Ethical considerations

• This dataset contains information about adverse events following vaccination
• Reports in VAERS do not establish causation
• Models trained on this data should not be used for medical decision-making without appropriate expert oversight
• Users should be aware of the limitations of passive surveillance data

License

This dataset is derived from U.S. government public data and is available under the Creative Commons Zero (CC0) license.

Citation

If you use this dataset, please cite the original VAERS data source:

Vaccine Adverse Event Reporting System (VAERS)
Centers for Disease Control and Prevention (CDC) and Food and Drug Administration (FDA)
Available at: https://vaers.hhs.gov/data.html

Dataset Creation

Created: 2025-11-10

Tool: VAERS Fine-Tuning Preprocessor Repository: https://github.com/chrisvoncsefalvay/vaers-ft-preprocessor

Processing configuration:

• Output field: SYMPTOM_TEXT
• Include history: No
• Years processed: 1990-2025
• Total examples: 2,649,322