cat-cog 0.2.0

Cognitive assessment and visual scoring powered by LLMs

cat-cog

Cognitive assessment and visual scoring powered by LLMs.

cat-cog provides LLM-powered evaluation of hand-drawn images for neuropsychological testing. It builds on cat-stack, the shared classification engine for the CatLLM ecosystem.

Installation

pip install cat-cog

This automatically installs cat-stack (the LLM classification engine).

CERAD Constructional Praxis Scoring

Score hand-drawn shapes (circle, diamond, overlapping rectangles, cube) using LLM vision models. The function sends images to the LLM, classifies drawing features, then applies CERAD scoring rules.

from cat_cog import cerad_drawn_score

# Score circle drawings (max score: 2)
results = cerad_drawn_score(
    shape="circle",
    image_input="./circle_drawings/",
    api_key=OPENAI_KEY,
    user_model="gpt-4o",
)
print(results[["image_file", "score"]])

Supported shapes

Shape	Max Score	Features Assessed
circle	2	Closure, circularity
diamond	3	4 sides, equal sides, resemblance
rectangles	2	Overlap, crossing pattern
cube	4	Front face, internal lines, parallel faces, 3D quality

Using with multiple models

All cat_stack.classify() parameters are available via **kwargs:

results = cerad_drawn_score(
    shape="cube",
    image_input="./cube_drawings/",
    api_key=OPENAI_KEY,
    models=[
        ("gpt-4o", "openai", OPENAI_KEY),
        ("claude-sonnet-4-20250514", "anthropic", ANTHROPIC_KEY),
    ],
)

Future

• Fine-tuned vision models for shape quality assessment (circle classifier, etc.)
• Additional cognitive screening instruments (clock drawing tests, etc.)

Ecosystem

Package	Domain
cat-stack	General-purpose classification engine (base)
cat-cog	Cognitive assessment & visual scoring (this package)
cat-survey	Survey response classification
cat-vader	Social media text
cat-ademic	Academic papers & PDFs
cat-pol	Political text

License

GPL-3.0-or-later