slicktune.recipes

Opinionated recipe helpers.

Submodules

• slicktune.recipes.probe

Classes

ProbeReport	Aggregate probe results.
ProbeResult	Outcome of a single probe question.

Functions

generate_reply(→ str)	Generate an assistant reply for prompt.
load_trained(→ tuple[Any, ...)	Load a saved adapter/model directory for probing.
prepare_model_for_inference(→ Any)	Switch a post-training model into a generate-safe state.
run_probes(→ ProbeReport)	Run probe questions against a model.

Package Contents

class slicktune.recipes.ProbeReport

Aggregate probe results.

Parameters:

results (list[ProbeResult]) – Per-question outcomes.

property pass_rate: float

Return fraction of probes that passed.

Returns:

float – Pass rate in [0, 1].

results: list[ProbeResult]

class slicktune.recipes.ProbeResult

Outcome of a single probe question.

Parameters:

• prompt (str) – User question.

• must_contain (str) – Expected substring (case-insensitive).

• generation (str) – Model completion.

• passed (bool) – Whether must_contain appears in generation.

generation: str

must_contain: str

passed: bool

prompt: str

slicktune.recipes.generate_reply(*, model: Any, tokenizer: transformers.PreTrainedTokenizerBase, prompt: str, max_new_tokens: int = 128, system_prompt: str = DEFAULT_SYSTEM_PROMPT) → str

Generate an assistant reply for prompt.

Parameters:

• model (Any) – Causal LM (base or PEFT).

• tokenizer (PreTrainedTokenizerBase) – Matching tokenizer.

• prompt (str) – User message.

• max_new_tokens (int, optional) – Generation length, by default 128.

• system_prompt (str, optional) – System message for chat-template models, by default the about-me prompt used for personal probes.

Returns:

str – Decoded assistant text.

slicktune.recipes.load_trained(output_dir: str | pathlib.Path) → tuple[Any, transformers.PreTrainedTokenizerBase]

Load a saved adapter/model directory for probing.

Parameters:

output_dir (str or Path) – Directory produced by Tuner.fit().

Returns:

tuple[Any, PreTrainedTokenizerBase] – (model, tokenizer).

slicktune.recipes.prepare_model_for_inference(model: Any) → Any

Switch a post-training model into a generate-safe state.

Disables gradient checkpointing (incompatible with KV cache) and enables use_cache so decoding does not collapse into token loops.

Parameters:

model (Any) – Trained base or PEFT model.

Returns:

Any – The same model, mutated for inference.

slicktune.recipes.run_probes(*, model: Any, tokenizer: transformers.PreTrainedTokenizerBase, probe_path: str | pathlib.Path, max_new_tokens: int = 128) → ProbeReport

Run probe questions against a model.

Parameters:

• model (Any) – Trained model.

• tokenizer (PreTrainedTokenizerBase) – Tokenizer.

• probe_path (str or Path) – JSONL probe file.

• max_new_tokens (int, optional) – Generation length, by default 128.

Returns:

ProbeReport – Aggregate probe outcomes.