How aimat-lab/llm-data-extraction works
Project type and purpose
• This is an academic “mthesis” repository for a Master’s thesis. It contains four main subsystems:
• expose: initial planning document (LaTeX)
• code: final experiment code, configs, and job scripts
• thesis_document: LaTeX source for the written thesis
• defense: LaTeX source and assets for the defense presentation
• The heart of the system is the code directory, which implements data loading, model definition, training, evaluation, and experiment orchestration.
File structure
• Top-level directories separate concerns:
• code/ – Python project (pyproject.toml, shell.nix, settings/, jobs/, src/)
• thesis_document/, expose/, defense/ – LaTeX projects for write-up and presentations
• Inside code/:
• settings/: YAML configs for different model sizes (7B, 13B, large)
• jobs/: batch job scripts (.job) that invoke the training pipeline on HPC
• src/mthesis/: Python package with modules:
– main.py (CLI entrypoint)
– dataloader.py (data ingestion)
– conversion.py (pre/post-processing)
– models.py (model instantiation)
– trainer.py (training loop orchestration)
– stats.py, confusion.py (evaluation metrics & confusion matrices)
– utils.py (shared utilities)
– train_example.py (demo or quick-start)
• Configuration: shell.nix fixes dependencies; pyproject.toml manages the Python package.