Skip to content

README.md

Latest commit

 

History

History
41 lines (30 loc) · 1.56 KB

README.md

File metadata and controls

41 lines (30 loc) · 1.56 KB

CastFactory

CastFactory is a recipe-centric research framework for LLM-driven time series forecasting. The current implementation provides the early framework scaffold for leakage-safe data handling, representation adapters, forecast parsing, evaluation, trace artifacts, and verifiable rewards.

Current Scope

Implemented foundation modules:

  • castfactory.core: recipe loading, experiment shell, and registries
  • castfactory.data: TSRecord, ForecastSample, CSV reader, timestamp split, window builder, scaler, leakage checks
  • castfactory.representation: statistics, context, textual summary, and hybrid representations
  • castfactory.parsers: JSON and array forecast parsers with fallback behavior
  • castfactory.evaluation: point metrics and standard evaluator
  • castfactory.trace: run artifact store
  • castfactory.rewards: accuracy, format, and composite rewards
  • castfactory.training: SFT dataset formatter, trainer shell, and optional Transformers backend

The architecture plan is in docs/CastFactory_Library_Architecture.md.

Development

python3 -m unittest discover -s tests -v
PYTHONPYCACHEPREFIX=/tmp/castfactory_pycache python3 -m compileall castfactory
python3 -m castfactory.cli.run /tmp/castfactory_cli_recipe.yaml

Optional dev tooling is declared in pyproject.toml:

python -m pip install -e ".[dev]"
python -m pytest tests -q
python -m ruff check castfactory tests

Optional training stacks:

python -m pip install -e ".[hf]"
python -m pip install -e ".[train]"