gepromptet/promptolution v2.3.0

Release v2.3.0

on GitHub

Release v2.3.0

What's changed

Promptolution v2.3.0 removes ExperimentConfig and its string dispatch get_* factories and adds config-driven runs and experiment grids via Hydra as well as a simple optimize() quickstart.

Added features:

• promptolution.experiment: config-driven single runs and whole experiment grids via the Hydra CLI, allowing for reproducible runs and research. When called through the CLI, Hydra builds a config from the nested config groups inside conf/ and the overrided that the caller provides. Running a sweep then becomes a CLI flag rather than a hand-written loop and a killed job resumes into the same output folder, skipping already completed experiment cells.
• promptolution.utils.dev_test_split and promptolution.utils.evaluate_prompts, split your data
  and score prompts on a task
• Optimizers generate initial_prompts from the task's task_description when none are given (moved here from the deleted promptolution.helpers.run_optimization).

Breaking changes:

• promptolution.helpers (run_experiment, run_optimization, run_evaluation, get_llm,
  get_task, get_optimizer, get_predictor, get_exemplar_selector) and ExperimentConfig are
  removed. Build components directly (see the README quickstart) or use experiment.
• Every component constructor drops its config parameter.
• hydra-core is now a main dependency; the [experiments] extra keeps the SLURM launcher plugin
  and HuggingFace datasets.

Full Changelog: here