Index

A

activation_fn (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute)

ACTPolicyWithAttention (class in src) (class in src.attention_maps.act_attention_mapper), [1]

B

batch_size (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute) block_size (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.sae.token_sampler.TokenSamplerConfig attribute) (src.sae.TokenSamplerConfig attribute) (src.SAETrainingConfig attribute) (src.TokenSamplerConfig attribute)

C

collect_activations() (src.sae.SAETrainer method) (src.sae.trainer.SAETrainer method) (src.SAETrainer method) compute_loss() (src.MultiModalSAE method) (src.sae.MultiModalSAE method) (src.sae.sae.MultiModalSAE method) create_default_ood_params_path() (in module src) (in module src.ood) (in module src.ood.ood_detector)

create_model() (src.sae.SAETrainer method) (src.sae.trainer.SAETrainer method) (src.SAETrainer method) create_optimizer_and_scheduler() (src.sae.SAETrainer method) (src.sae.trainer.SAETrainer method) (src.SAETrainer method)

D

decode() (src.MultiModalSAE method) (src.sae.MultiModalSAE method) (src.sae.sae.MultiModalSAE method)

device (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute)

E

early_stopping_min_delta (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute) early_stopping_patience (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute)

encode() (src.MultiModalSAE method) (src.sae.MultiModalSAE method) (src.sae.sae.MultiModalSAE method) expansion_factor (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute)

F

feature_dim (src.sae.config.SAETrainingConfig property) (src.sae.SAETrainingConfig property) (src.SAETrainingConfig property) fit_ood_threshold_to_validation_dataset() (src.ood.ood_detector.OODDetector method) (src.ood.OODDetector method) (src.OODDetector method) fixed_tokens (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.sae.token_sampler.TokenSamplerConfig attribute) (src.sae.TokenSamplerConfig attribute) (src.SAETrainingConfig attribute) (src.TokenSamplerConfig attribute)

forward() (src.MultiModalSAE method) (src.sae.MultiModalSAE method) (src.sae.sae.MultiModalSAE method) from_default_path() (src.sae.builder.SAEBuilder class method) (src.sae.SAEBuilder class method) (src.SAEBuilder class method)

G

generate_model_card() (src.sae.SAETrainer method) (src.sae.trainer.SAETrainer method) (src.SAETrainer method) get_ood_stats() (src.ood.ood_detector.OODDetector method) (src.ood.OODDetector method) (src.OODDetector method) get_reconstruction_error() (src.ood.ood_detector.OODDetector method) (src.ood.OODDetector method) (src.OODDetector method)

get_sampling_info() (src.sae.token_sampler.TokenSampler method) (src.sae.TokenSampler method) (src.TokenSampler method) gradient_clip_norm (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute)

I

infer_model_params_from_cache() (src.sae.config.SAETrainingConfig method) (src.sae.SAETrainingConfig method) (src.SAETrainingConfig method)

is_out_of_distribution() (src.ood.ood_detector.OODDetector method) (src.ood.OODDetector method) (src.OODDetector method)

L

l1_penalty (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute) learning_rate (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute) load_checkpoint() (src.sae.SAETrainer method) (src.sae.trainer.SAETrainer method) (src.SAETrainer method) load_from_experiment() (src.sae.builder.SAEBuilder method) (src.sae.SAEBuilder method) (src.SAEBuilder method) load_from_files() (src.sae.builder.SAEBuilder method) (src.sae.SAEBuilder method) (src.SAEBuilder method) load_from_hub() (src.sae.builder.SAEBuilder method) (src.sae.SAEBuilder method) (src.SAEBuilder method)

load_original_num_tokens_from_cache() (in module src.sae) load_sae_from_hub() (in module src.sae.trainer) load_sae_model() (in module src) (in module src.sae) (in module src.sae.trainer) load_sae_model_simple() (in module src) (in module src.sae) (in module src.sae.builder) load_with_auto_config() (src.sae.builder.SAEBuilder method) (src.sae.SAEBuilder method) (src.SAEBuilder method) log_every (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute) lr_schedule (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute)

M

make_dataset_without_config() (in module src) (in module src.utils.dataset_factory), [1] max_sampled_tokens (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.sae.token_sampler.TokenSamplerConfig attribute) (src.sae.TokenSamplerConfig attribute) (src.SAETrainingConfig attribute) (src.TokenSamplerConfig attribute) module src src.attention_maps src.attention_maps.act_attention_mapper, [1] src.ood src.ood.ood_detector src.sae src.sae.builder src.sae.config src.sae.sae src.sae.token_sampler src.sae.trainer src.utils src.utils.dataset_factory, [1]

MultiModalSAE (class in src) (class in src.sae) (class in src.sae.sae)

N

needs_ood_fitting() (src.ood.ood_detector.OODDetector method) (src.ood.OODDetector method) (src.OODDetector method) num_epochs (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute)

num_tokens (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute)

O

OODDetector (class in src) (class in src.ood) (class in src.ood.ood_detector)

optimizer (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute)

P

push_model_to_hub() (src.sae.SAETrainer method) (src.sae.trainer.SAETrainer method) (src.SAETrainer method)

R

random_seed (src.sae.token_sampler.TokenSamplerConfig attribute) (src.sae.TokenSamplerConfig attribute) (src.TokenSamplerConfig attribute)

resolve_delta_timestamps_without_config() (in module src.utils.dataset_factory), [1]

S

SAEBuilder (class in src) (class in src.sae) (class in src.sae.builder) SAETrainer (class in src) (class in src.sae) (class in src.sae.trainer) SAETrainingConfig (class in src) (class in src.sae) (class in src.sae.config) sample_tokens() (src.sae.token_sampler.TokenSampler method) (src.sae.TokenSampler method) (src.TokenSampler method) sampling_strategy (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.sae.token_sampler.TokenSamplerConfig attribute) (src.sae.TokenSamplerConfig attribute) (src.SAETrainingConfig attribute) (src.TokenSamplerConfig attribute) sampling_stride (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.sae.token_sampler.TokenSamplerConfig attribute) (src.sae.TokenSamplerConfig attribute) (src.SAETrainingConfig attribute) (src.TokenSamplerConfig attribute) save_checkpoint() (src.sae.SAETrainer method) (src.sae.trainer.SAETrainer method) (src.SAETrainer method) save_complete_model() (src.sae.SAETrainer method) (src.sae.trainer.SAETrainer method) (src.SAETrainer method)

save_every (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute) select_action() (src.ACTPolicyWithAttention method) (src.attention_maps.act_attention_mapper.ACTPolicyWithAttention method), [1] src module src.attention_maps module src.attention_maps.act_attention_mapper module, [1] src.ood module src.ood.ood_detector module src.sae module src.sae.builder module src.sae.config module src.sae.sae module src.sae.token_sampler module src.sae.trainer module src.utils module src.utils.dataset_factory module, [1]

T

token_dim (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute) TokenSampler (class in src) (class in src.sae) (class in src.sae.token_sampler) TokenSamplerConfig (class in src) (class in src.sae) (class in src.sae.token_sampler)

train() (src.sae.SAETrainer method) (src.sae.trainer.SAETrainer method) (src.SAETrainer method) train_step() (src.sae.SAETrainer method) (src.sae.trainer.SAETrainer method) (src.SAETrainer method)

U

use_token_sampling (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute)

V

validate_every (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute) validation_split (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute)

visualize_attention() (src.ACTPolicyWithAttention method) (src.attention_maps.act_attention_mapper.ACTPolicyWithAttention method), [1]

W

warmup_epochs (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute)

weight_decay (src.sae.config.SAETrainingConfig attribute) (src.sae.SAETrainingConfig attribute) (src.SAETrainingConfig attribute)