slicktune.cli.main

Click CLI for slicktune.

Attributes

StrategyName
console

Functions

cli(→ None)	slicktune: composable LLM fine-tuning.
eval_cmd(→ None)	Run holdout perplexity and/or probe judging on a saved checkpoint.
merge_cmd(→ None)	Combine PEFT adapters with TIES / DARE / linear-style merges.
probe(→ None)	Ask probe questions to verify personal facts were learned.
train(→ None)	Fine-tune with the chosen objective + parameter strategy.

Module Contents

slicktune.cli.main.StrategyName

slicktune.cli.main.cli() → None

slicktune: composable LLM fine-tuning.

slicktune.cli.main.console

slicktune.cli.main.eval_cmd(model_dir: pathlib.Path, eval_data: pathlib.Path | None, probe_path: pathlib.Path | None, judge_name: str, max_seq_length: int, max_new_tokens: int) → None

Run holdout perplexity and/or probe judging on a saved checkpoint.

slicktune.cli.main.merge_cmd(model_id: str, adapter_specs: tuple[str, Ellipsis], method: str, density: float, output_dir: pathlib.Path, bake: bool, combined_name: str) → None

Combine PEFT adapters with TIES / DARE / linear-style merges.

slicktune.cli.main.probe(model_dir: pathlib.Path, probe_path: pathlib.Path, max_new_tokens: int) → None

Ask probe questions to verify personal facts were learned.

slicktune.cli.main.train(model_id: str, strategy: str, objective_name: str, beta: float, num_generations: int, max_completion_length: int, data_path: pathlib.Path, eval_data: pathlib.Path | None, probe_path: pathlib.Path | None, output_dir: pathlib.Path, epochs: float, lr: float, max_seq_length: int, batch_size: int, grad_accum: int) → None

Fine-tune with the chosen objective + parameter strategy.