aquin.compute.steer_vector Namespace Reference

Functions
str	_now_iso ()
Path	_experiment_path (str model_id)
dict[str, Any]|None	lookup_probe_record (str model_id, str|None probe_id=None)
tuple[torch.Tensor, int, dict[str, Any]]	resolve_feature_direction (str model_id, int feature_idx, *, int|None layer=None, bool apply_norm_std=True)
dict[str, Any]	build_steer_vector_payload (str model_id, int feature_idx, *, int|None layer=None, str|None feature_label=None, str|None probe_id=None)
Path	write_steer_vector_file (dict[str, Any] payload, str|Path output)
dict[str, Any]	load_steer_vector_file (str|Path path)
torch.Tensor	steer_vector_tensor (dict[str, Any] data, *, torch.device|str device, torch.dtype dtype)
dict[str, Any]	extract_steer_vector (str model_id, int feature_idx, str|Path output, *, int|None layer=None, str|None feature_label=None, str|None probe_id=None)
str	_generate_steered_response (Any model, str prompt, *, torch.Tensor steer_vec, float steer_strength, int layer, int max_new_tokens)
dict[str, Any]	run_steer_with_vector (*, str model_id, str|None prompt=None, float steer_strength, int|None layer=None, int|None feature_idx=None, str|Path|None vector_path=None, dict[str, Any]|None vector_data=None, str|None feature_label=None, int max_new_tokens=80, dict|None ctx=None, dict|None args=None)

Variables
int	SCHEMA_VERSION = 1
str	KIND = "sae_decoder_steer"

Function Documentation

_experiment_path()

Path _experiment_path ( str model_id )

protected

Definition at line 24 of file steer_vector.py.

Referenced by lookup_probe_record().

_generate_steered_response()

str _generate_steered_response ( Any model, str prompt, * , torch.Tensor steer_vec, float steer_strength, int layer, int max_new_tokens )

protected

Definition at line 203 of file steer_vector.py.

Referenced by run_steer_with_vector().

_now_iso()

str _now_iso ( )

protected

Definition at line 20 of file steer_vector.py.

Referenced by build_steer_vector_payload().

build_steer_vector_payload()

dict[str, Any] build_steer_vector_payload	(	str	model_id,
		int	feature_idx,
		*	,
		int | None	layer = None,
		str | None	feature_label = None,
		str | None	probe_id = None )

Definition at line 101 of file steer_vector.py.

References _now_iso(), lookup_probe_record(), and resolve_feature_direction().

Referenced by extract_steer_vector().

extract_steer_vector()

dict[str, Any] extract_steer_vector	(	str	model_id,
		int	feature_idx,
		str | Path	output,
		*	,
		int | None	layer = None,
		str | None	feature_label = None,
		str | None	probe_id = None )

Definition at line 169 of file steer_vector.py.

References build_steer_vector_payload(), and write_steer_vector_file().

load_steer_vector_file()

dict[str, Any] load_steer_vector_file

(

str | Path

path

)

Definition at line 140 of file steer_vector.py.

Referenced by run_steer_with_vector().

lookup_probe_record()

dict[str, Any] | None lookup_probe_record	(	str	model_id,
		str | None	probe_id = None )

Load persisted find-feature record from ~/.aquin/experiments/<model>.json.

Definition at line 29 of file steer_vector.py.

References _experiment_path().

Referenced by build_steer_vector_payload().

resolve_feature_direction()

tuple[torch.Tensor, int, dict[str, Any]] resolve_feature_direction	(	str	model_id,
		int	feature_idx,
		*	,
		int | None	layer = None,
		bool	apply_norm_std = True )

Return unit-ready decoder direction (d_model,) and resolved layer.

Definition at line 45 of file steer_vector.py.

Referenced by build_steer_vector_payload(), and run_steer_with_vector().

run_steer_with_vector()

dict[str, Any] run_steer_with_vector	(	*	,
		str	model_id,
		str | None	prompt = None,
		float	steer_strength,
		int | None	layer = None,
		int | None	feature_idx = None,
		str | Path | None	vector_path = None,
		dict[str, Any] | None	vector_data = None,
		str | None	feature_label = None,
		int	max_new_tokens = 80,
		dict | None	ctx = None,
		dict | None	args = None )

Steer generation using a live feature index or a saved vector file.

Definition at line 234 of file steer_vector.py.

References _generate_steered_response(), load_steer_vector_file(), resolve_feature_direction(), and steer_vector_tensor().

steer_vector_tensor()

torch.Tensor steer_vector_tensor	(	dict[str, Any]	data,
		*	,
		torch.device | str	device,
		torch.dtype	dtype )

Definition at line 157 of file steer_vector.py.

Referenced by run_steer_with_vector().

write_steer_vector_file()

Path write_steer_vector_file	(	dict[str, Any]	payload,
		str | Path	output )

Definition at line 133 of file steer_vector.py.

Referenced by extract_steer_vector().

Variable Documentation

KIND

str aquin.compute.steer_vector.KIND = "sae_decoder_steer"

Definition at line 17 of file steer_vector.py.

SCHEMA_VERSION

int aquin.compute.steer_vector.SCHEMA_VERSION = 1

Definition at line 16 of file steer_vector.py.