Skip to content

CLAUDE.md

Latest commit

 

History

History
64 lines (41 loc) · 2.39 KB

CLAUDE.md

File metadata and controls

64 lines (41 loc) · 2.39 KB
Raw

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

openllm-selector is a Python tool for researchers to filter, compare, and select open Large Language Models (LLMs) based on characteristics like size, open weights, checkpoints, and modality. It is in early development (v0.0.1).

Setup

pip install -e '.[dev]'

This installs the package in editable mode with dev dependencies: pytest, streamlit, plotly, requests.

Common Commands

# Run tests
pytest

# Run a single test file
pytest tests/test_example.py

# Run a single test
pytest tests/test_example.py::test_function_name

Architecture

The package source lives in src/openllm_selector/ (src-layout). The __init__.py re-exports everything from submodules via from .example import *.

Planned features include:

  • A curated LLM database with filtering/comparison
  • Access to foundational and recent papers per model
  • Visualization of model clusters (likely via Streamlit + Plotly)

The dev extras (streamlit, plotly, requests) suggest the intended UI will be a Streamlit app consuming data via HTTP requests.

Build

pip install hatchling
python -m hatchling build

Uses hatchling as the build backend. The wheel target is configured to package openllm_selector from the src/ directory.

Rules

Git: Never merge branches or create pull requests. Commit and push only.

openness_score: This field is computed dynamically inside load_models() by compute_openness_score() and injected into each record at load time. It must never be stored in models.json.

Model data: All values in models.json were manually verified against primary sources (HuggingFace model pages and foundational papers). Any change to model data must be verified against the same primary sources before committing.

App data access: Model data must never be duplicated or hardcoded in app code. All model data must be read from the database through the caching wrappers in app/utils.py (e.g. cached_load_models(), cached_get_families()).

AI Usage

Claude Code was used to assist with code generation, documentation, and unit tests. Prompts followed a consistent pattern: specify the function name, its parameters, expected behaviour, and edge cases to test. All generated code was reviewed and tested before committing.