Tools for extracting and comparing political committee names from fundraising emails using local LLMs. Given a JSON file of parsed political emails, these scripts use small language models (via LLM) to identify the sponsoring committee from each email’s disclaimer, then compare results across models using fuzzy matching.
The source data file emails.json is too large for version control. Download it from Google Drive and place it in the project root before running.
Each record in emails.json contains fields like email, date, subject, body, party, and disclaimer.
Requires uv and Python 3.12+.
git clone https://github.com/dwillis/political-fundraising-emails.git
cd political-fundraising-emails
uv sync
You also need Ollama (or another LLM backend supported by the llm CLI) with the models you want to use pulled locally.
Runs a local LLM against each email’s body text to extract the committee name from political disclaimers. Processes emails in parallel and supports resuming interrupted runs.
# Process all emails using the default model (qwen3.5:4b)
uv run python email_add_cmte.py emails.json
# Use 8 workers
uv run python email_add_cmte.py emails.json --workers 8
# Test with first 10 emails
uv run python email_add_cmte.py emails.json --test 10
Output is written to emails_updated.json (derived from the input filename). A .progress.jsonl file tracks incremental results so interrupted runs can resume.
To change the model, edit the model_name variable at the top of the script. The model must be available through your llm installation.
Compares committee extractions across multiple model output files. Reports exact and fuzzy match rates, NA distributions, and samples of disagreements.
# Compare three model outputs (defaults)
uv run python compare_committees.py
# Specify custom file paths
uv run python compare_committees.py --gemma results_gemma.json --qwen35 results_qwen35.json --qwen3 results_qwen3.json
# Only compare emails with disclaimers
uv run python compare_committees.py --disclaimer
# Export results
uv run python compare_committees.py --export-json stats.json --export-csv disagreements.csv
MIT License - see LICENSE for details.
Derek Willis