cat-pol 1.2.0

Policy document classification powered by LLMs

cat-pol

Political text classification and analysis powered by LLMs. A policy-specific wrapper around cat-stack with built-in access to 15 political data sources on HuggingFace.

Installation

pip install cat-pol

With optional extras:

pip install "cat-pol[pdf]"         # PDF document processing
pip install "cat-pol[embeddings]"  # Embedding-based similarity scoring
pip install "cat-pol[sources]"     # Data source loading (datasets, huggingface_hub)

Quick Start

Classify ordinances from a built-in source

import cat_pol as pol

results = pol.classify(
    source="city_san_diego",
    categories=["Housing", "Public Safety", "Infrastructure", "Finance"],
    doc_type="ordinance",
    since="2022-01-01",
    n=50,
    api_key="sk-...",
)

Classify raw text

results = pol.classify(
    input_data=[
        "The committee voted to approve the rezoning request for parcel 42.",
        "Motion to table the budget amendment until the next session.",
    ],
    categories=["Approval", "Rejection", "Deferral", "Amendment"],
    document_context="City council meeting minutes",
    api_key="sk-...",
)

Optimize prompts with user feedback

result = pol.prompt_tune(
    source="city_san_diego",
    categories=["Pro-Business", "Pro-Regulation", "Tax Increase", "Tax Decrease"],
    doc_type="ordinance",
    since="2020-01-01",
    n=100,
    api_key="sk-...",
    sample_size=15,
)

# Use the optimized prompt for full classification
results = pol.classify(
    source="city_san_diego",
    categories=["Pro-Business", "Pro-Regulation", "Tax Increase", "Tax Decrease"],
    system_prompt=result["system_prompt"],
    api_key="sk-...",
)

Summarize with different formats

# Bullet points
pol.summarize(source="federal_executive_orders", n=10, format="bullets", api_key="sk-...")

# Full report
pol.summarize(source="federal_laws", n=5, format="report", api_key="sk-...")

# One-liner
pol.summarize(source="social_trump_truth", since="2024-01-01", n=20, format="one-liner", api_key="sk-...")

Discover categories

result = pol.extract(
    source="city_berkeley",
    n=200,
    api_key="sk-...",
)
print(result["top_categories"])

Fetch raw data

# List all sources
pol.list_sources()
pol.list_sources(level="city")
pol.list_sources(level="federal")

# Fetch data
df = pol.fetch_source("city_san_diego", n=100, since="2020-01-01", doc_type="ordinance")
df = pol.fetch_source("federal_executive_orders", n=50)
df = pol.fetch_source("social_trump_truth", since="2024-01-01")

Data Sources

All datasets are public on HuggingFace — no authentication required.

California Cities

Source	Rows	Types	Repo
city_san_diego	87,983	ordinances, resolutions	chrissoria/san-diego-ordinances
city_los_angeles	34,427	ordinances	chrissoria/la-ordinances
city_berkeley	9,028	ordinances	chrissoria/berkeley-ordinances
city_san_francisco	4,033	ordinances	chrissoria/sf-ordinances
city_long_beach	3,898	ordinances, resolutions	chrissoria/long-beach-ordinances
city_bakersfield	2,655	ordinances	chrissoria/bakersfield-ordinances
city_newport_beach	2,719	ordinances	chrissoria/newport-beach-ordinances
city_salinas	2,574	ordinances, resolutions	chrissoria/salinas-ordinances
city_clovis	2,343	ordinances	chrissoria/clovis-ordinances
city_oakland	1,824	ordinances	chrissoria/oakland-ordinances
city_fresno	706	ordinances, resolutions	chrissoria/fresno-ordinances

Federal

Source	Rows	Types	Repo
federal_laws	5,915	public laws (1995–present)	chrissoria/federal-public-laws
federal_executive_orders	1,530+	executive orders	chrissoria/executive-orders
federal_speeches	305	SOTU, inaugurals	chrissoria/presidential-speeches

Social Media

Source	Rows	Types	Repo
social_trump_truth	32,000+	Truth Social posts	chrissoria/trump-truth-social

All sources are updated weekly (Sundays at 9 AM) via automated scrapers. Truth Social is updated daily at 9 AM.

Trump Truth Social Dataset Columns

The social_trump_truth dataset is enriched with metadata, market data, and image descriptions:

Post metadata:

Column	Description
date	Post date (YYYY-MM-DD)
time	Post time in UTC (HH:MM:SS)
day_of_week	Day name (Monday, Tuesday, etc.)
datetime	Full ISO timestamp
text	Post text content
url	Truth Social post URL
post_id	Unique post identifier
is_president	Whether Trump was president at time of post
is_president_elect	Whether Trump was president-elect at time of post
replies_count	Number of replies
reblogs_count	Number of reblogs
favourites_count	Number of favourites
media_urls	Image/video URLs attached to the post
has_media	Whether the post has media attachments
image_alt_text	AI-generated factual image description (alt-text format)

Market data (18 tickers):

Each ticker has 7 columns following the convention {ticker}_{metric}:

Metric	Description
{ticker}_open	Daily open price
{ticker}_close	Daily close price
{ticker}_1hr_before	Price 1 hour before the post
{ticker}_5min_before	Price 5 minutes before the post
{ticker}_at_post	Price at time of post
{ticker}_5min_after	Price 5 minutes after the post
{ticker}_1hr_after	Price 1 hour after the post

Tickers included:

Ticker	Name	Category
sp500	S&P 500 (^GSPC)	Broad market
dia	SPDR Dow Jones Industrial Average ETF	Broad market
qqq	Invesco QQQ (Nasdaq-100)	Tech/growth
djt	Trump Media & Technology Group	Trump-linked
lmt	Lockheed Martin	Defense
war	Themes US Military Academy ETF	Defense
xli	Industrial Select Sector SPDR	Industrials
xlv	Health Care Select Sector SPDR	Healthcare
xph	SPDR S&P Pharmaceuticals ETF	Pharma
cnrg	SPDR S&P Kensho Clean Power ETF	Clean energy
gld	SPDR Gold Shares	Gold/commodities
uso	United States Oil Fund	Oil/energy
fxi	iShares China Large-Cap ETF	China/trade
eww	iShares MSCI Mexico ETF	Mexico/trade
vgk	Vanguard FTSE Europe ETF	Europe
ibit	iShares Bitcoin ETF	Crypto
tlt	iShares 20+ Year Treasury Bond ETF	Bonds/rates
uup	Invesco DB US Dollar Index	USD strength

Intraday prices use the highest available resolution: 1-minute (last ~7 days), 5-minute (last ~60 days), or hourly (last ~2 years). Weekend/holiday posts use the most recent trading day's values. The sp500_resolution column indicates the data resolution used.

API

Function	Description
classify()	Classify text into predefined categories
prompt_tune()	Optimize classification prompts via user feedback
extract()	Discover and normalize categories from text
explore()	Raw category extraction (no deduplication)
summarize()	Summarize text, PDFs, or image URLs with format options (paragraph, bullets, one-liner, structured, report, alt-text)
list_sources()	List available data sources
fetch_source()	Fetch raw data from a source

All functions accept either input_data= (raw text, files, directories) or source= (pull from HuggingFace). All cat-stack parameters (multi-model ensemble, batch mode, chain-of-thought, etc.) pass through via **kwargs.

Ecosystem

Package	Role
cat-stack	Domain-agnostic LLM classification engine
cat-pol	Political text classification (this package)
cat-vader	Social media text (Reddit, Twitter/X)
cat-ademic	Academic papers and citations

License

GPL-3.0-or-later