aquin.compute.find_feature Namespace Reference

Functions
str	normalize_direction (str|None direction)
Path	_resolve_deception_fixture ()
str	_row_text (dict[str, Any] row, *str keys)
tuple[list[str], list[str], dict[str, Any]]	load_deception_probes (str|Path|None path)
list[dict[str, Any]]	_rank_features (torch.Tensor honest_acts, torch.Tensor deceptive_acts, *, int top_k, str direction="both")
str|None	_find_feature_warning (str direction, dict[str, Any]|None chosen, *, bool persisted, str conditioning="behavior", dict[str, Any]|None behavior_meta=None)
str	persist_experiment_record (str model_id, str key, dict[str, Any] record, *, str|None session_id=None)
dict[str, Any]	run_find_feature (str model_id, *, str scorer="deception", str|Path|None prompts_path=None, int|None layer=None, str|Path|None checkpoint_path=None, int top_k=20, str direction="both", str conditioning="behavior", int max_new_tokens=64, float temperature=0.0, int benchmark_top=0, str|None persist_key=None, str|None session_id=None, Any|None openai_client=None)

Variables
	HONEST_LABELS = frozenset({"honest", "truthful", "true", "control", "baseline"})
	DECEPTIVE_LABELS = frozenset({"deceptive", "deception", "lie", "lying", "dishonest", "false"})
tuple	TEXT_KEYS = ("instruction", "prompt", "text", "content", "honest", "deceptive")
	VALID_DIRECTIONS = frozenset({"both", "deceptive", "honest"})

Function Documentation

_find_feature_warning()

str | None _find_feature_warning ( str direction, dict[str, Any] | None chosen, * , bool persisted, str conditioning = "behavior", dict[str, Any] | None behavior_meta = None )

protected

Definition at line 162 of file find_feature.py.

References normalize_direction().

Referenced by run_find_feature().

_rank_features()

list[dict[str, Any]] _rank_features ( torch.Tensor honest_acts, torch.Tensor deceptive_acts, * , int top_k, str direction = "both" )

protected

honest_acts, deceptive_acts: (n_features,) mean activations.

Definition at line 112 of file find_feature.py.

References normalize_direction().

Referenced by run_find_feature().

_resolve_deception_fixture()

Path _resolve_deception_fixture ( )

protected

Bundled probes were removed. Callers must pass prompts=.

Definition at line 32 of file find_feature.py.

Referenced by load_deception_probes().

_row_text()

str _row_text ( dict[str, Any] row, *str keys )

protected

Definition at line 39 of file find_feature.py.

Referenced by load_deception_probes().

load_deception_probes()

tuple[list[str], list[str], dict[str, Any]] load_deception_probes

(

str | Path | None

path

)

Load honest and deceptive text lists from JSON/JSONL.

Supports:
  - paired rows: honest + deceptive fields on one object
  - labeled rows: instruction/prompt/text + label honest|deceptive

Definition at line 47 of file find_feature.py.

References _resolve_deception_fixture(), and _row_text().

Referenced by run_find_feature().

normalize_direction()

str normalize_direction

(

str | None

direction

)

Definition at line 25 of file find_feature.py.

Referenced by _find_feature_warning(), _rank_features(), and run_find_feature().

persist_experiment_record()

str persist_experiment_record	(	str	model_id,
		str	key,
		dict[str, Any]	record,
		*	,
		str | None	session_id = None )

Write canonical feature record to ~/.aquin/experiments/<model>.json and optional session memory.

Definition at line 225 of file find_feature.py.

Referenced by run_find_feature().

run_find_feature()

dict[str, Any] run_find_feature	(	str	model_id,
		*	,
		str	scorer = "deception",
		str | Path | None	prompts_path = None,
		int | None	layer = None,
		str | Path | None	checkpoint_path = None,
		int	top_k = 20,
		str	direction = "both",
		str	conditioning = "behavior",
		int	max_new_tokens = 64,
		float	temperature = 0.0,
		int	benchmark_top = 0,
		str | None	persist_key = None,
		str | None	session_id = None,
		Any | None	openai_client = None )

Definition at line 258 of file find_feature.py.

References _find_feature_warning(), _rank_features(), load_deception_probes(), normalize_direction(), and persist_experiment_record().

Variable Documentation

DECEPTIVE_LABELS

aquin.compute.find_feature.DECEPTIVE_LABELS = frozenset({"deceptive", "deception", "lie", "lying", "dishonest", "false"})

Definition at line 20 of file find_feature.py.

HONEST_LABELS

aquin.compute.find_feature.HONEST_LABELS = frozenset({"honest", "truthful", "true", "control", "baseline"})

Definition at line 19 of file find_feature.py.

TEXT_KEYS

tuple aquin.compute.find_feature.TEXT_KEYS = ("instruction", "prompt", "text", "content", "honest", "deceptive")

Definition at line 21 of file find_feature.py.

VALID_DIRECTIONS

aquin.compute.find_feature.VALID_DIRECTIONS = frozenset({"both", "deceptive", "honest"})

Definition at line 22 of file find_feature.py.