GraphLM

한국어 | English (TBD)

Research codebase exploring training-time dynamic parameter count for Transformers — graph representation is the tool, dynamic growth / pruning / architecture search during training is the goal. (Not sparse activation like MoE: those keep total params fixed.)

Python 3.11+, Jupyter notebooks for experiments, package code in src/graphlm/.

See CLAUDE.md for the full paradigm definition.

Quick start

pip install -e ".[dev]"
make test
make lint

See docs/ko/README.md for the full guide, and .claude/rules/ for development conventions.

Project layout

src/graphlm/    Python package (models, graph utils, training, eval)
notebooks/      Jupyter experiment notebooks
tests/          pytest tests (mirrors src/graphlm/)
data/           experiment data (gitignored)
docs/           CI docs, Korean documentation, paper summaries (docs/papers/)
.claude/        AI collaboration rules and infra

License

TBD.