Upload 40 files

l1_coeff=2_expansion=0.5_control=0.0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec79ba0e38de40717b76f79b006f1d37d36403587453c531c22d0dd9d2ca32fc
+size 78015904

l1_coeff=2_expansion=0.5_control=0.0_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": null, "is_control_dataset_tokenized": true, "control_mixture": 0.0, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 384, "b_dec_init_method": "zeros", "expansion_factor": 0.5, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 2, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=2_expansion=0.5_control_mix=0.0_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=2_expansion=0.5_tokens=24576000_lr=0.001/2ufmaml4", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=2_expansion=0.5_control=0.1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:36e2f5095a71b9ee6e9c16ed434acd1262d682cee0b70bc5839d0840720b4f22
+size 78015904

l1_coeff=2_expansion=0.5_control=0.1_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": "NeelNanda/openwebtext-tokenized-9b", "is_control_dataset_tokenized": true, "control_mixture": 0.1, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 384, "b_dec_init_method": "zeros", "expansion_factor": 0.5, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 2, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=2_expansion=0.5_control_mix=0.1_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=2_expansion=0.5_tokens=24576000_lr=0.001/ic4wpyr5", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=2_expansion=2_control=0.0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3c6981a1f7c49124439e96bdecfcc8b2389cad3d39f42b60807e2c17eeeb9f3
+size 85107624

l1_coeff=2_expansion=2_control=0.0_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": null, "is_control_dataset_tokenized": true, "control_mixture": 0.0, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 1536, "b_dec_init_method": "zeros", "expansion_factor": 2, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 2, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=2_expansion=2_control_mix=0.0_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=2_expansion=2_tokens=24576000_lr=0.001/6m79e4gv", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=2_expansion=2_control=0.1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:daf7d58c01cd3c5082714cceadb528eba30a531965234d0370d47733484e93d4
+size 85107624

l1_coeff=2_expansion=2_control=0.1_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": "NeelNanda/openwebtext-tokenized-9b", "is_control_dataset_tokenized": true, "control_mixture": 0.1, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 1536, "b_dec_init_method": "zeros", "expansion_factor": 2, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 2, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=2_expansion=2_control_mix=0.1_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=2_expansion=2_tokens=24576000_lr=0.001/q07gu3q8", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=3_expansion=0.5_control=0.0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd63d1acce0edceff53dadcaf3ac9dbe75f042525a61e46892f5ea28008fda59
+size 78015904

l1_coeff=3_expansion=0.5_control=0.0_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": null, "is_control_dataset_tokenized": true, "control_mixture": 0.0, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 384, "b_dec_init_method": "zeros", "expansion_factor": 0.5, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 3, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=3_expansion=0.5_control_mix=0.0_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=3_expansion=0.5_tokens=24576000_lr=0.001/v1cvmyl0", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=3_expansion=0.5_control=0.1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c9726f19f13cd4142680d937213508121d145b1d1752774be8f0a55a2f185d3d
+size 78015904

l1_coeff=3_expansion=0.5_control=0.1_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": "NeelNanda/openwebtext-tokenized-9b", "is_control_dataset_tokenized": true, "control_mixture": 0.1, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 384, "b_dec_init_method": "zeros", "expansion_factor": 0.5, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 3, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=3_expansion=0.5_control_mix=0.1_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=3_expansion=0.5_tokens=24576000_lr=0.001/3u0gw6op", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=3_expansion=2_control=0.0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d291c37fefdd658ed1fa2f0951407fcbc5fc9c788c7926359f06d129a2855ca
+size 85107624

l1_coeff=3_expansion=2_control=0.0_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": null, "is_control_dataset_tokenized": true, "control_mixture": 0.0, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 1536, "b_dec_init_method": "zeros", "expansion_factor": 2, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 3, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=3_expansion=2_control_mix=0.0_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=3_expansion=2_tokens=24576000_lr=0.001/sv8g8jbs", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=3_expansion=2_control=0.1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:80e5fc5e0e5a37cb2d0eda70c0dd1b101e2c72e01de2848d0be10d673498e72b
+size 85107624

l1_coeff=3_expansion=2_control=0.1_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": "NeelNanda/openwebtext-tokenized-9b", "is_control_dataset_tokenized": true, "control_mixture": 0.1, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 1536, "b_dec_init_method": "zeros", "expansion_factor": 2, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 3, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=3_expansion=2_control_mix=0.1_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=3_expansion=2_tokens=24576000_lr=0.001/i1ywk17b", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=4_expansion=0.5_control=0.0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:91c4f61bb288ceb0c35d248e8e24181eda7b5c84c70da138ad8e554111de4279
+size 78015904

l1_coeff=4_expansion=0.5_control=0.0_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": null, "is_control_dataset_tokenized": true, "control_mixture": 0.0, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 384, "b_dec_init_method": "zeros", "expansion_factor": 0.5, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 4, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=4_expansion=0.5_control_mix=0.0_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=4_expansion=0.5_tokens=24576000_lr=0.001/1hub44j5", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=4_expansion=0.5_control=0.1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b8f2d2509f9705658ed1ebebb056cd31d80fe8c4dd8604903aa4d59dd06b7c3
+size 78015904

l1_coeff=4_expansion=0.5_control=0.1_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": "NeelNanda/openwebtext-tokenized-9b", "is_control_dataset_tokenized": true, "control_mixture": 0.1, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 384, "b_dec_init_method": "zeros", "expansion_factor": 0.5, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 4, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=4_expansion=0.5_control_mix=0.1_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=4_expansion=0.5_tokens=24576000_lr=0.001/9xst9se3", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=4_expansion=2_control=0.0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:467dbb50f3992a06819cded017b37846e29cc7bdf8c4bba74673a0b8fd10ee60
+size 85107624

l1_coeff=4_expansion=2_control=0.0_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": null, "is_control_dataset_tokenized": true, "control_mixture": 0.0, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 1536, "b_dec_init_method": "zeros", "expansion_factor": 2, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 4, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=4_expansion=2_control_mix=0.0_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=4_expansion=2_tokens=24576000_lr=0.001/2l3446hj", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=4_expansion=2_control=0.1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59a1ed7574085f895b30997c67f4024eb3f0a76d38305272f0068c22b8df17b5
+size 85107624

l1_coeff=4_expansion=2_control=0.1_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": "NeelNanda/openwebtext-tokenized-9b", "is_control_dataset_tokenized": true, "control_mixture": 0.1, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 1536, "b_dec_init_method": "zeros", "expansion_factor": 2, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 4, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=4_expansion=2_control_mix=0.1_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=4_expansion=2_tokens=24576000_lr=0.001/1jcg7559", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=5_expansion=0.5_control=0.0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8122db9e5d928ed9e2fbf97dfef8e108beb228cf0857f2698e478da5b50b4c6b
+size 78015904

l1_coeff=5_expansion=0.5_control=0.0_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": null, "is_control_dataset_tokenized": true, "control_mixture": 0.0, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 384, "b_dec_init_method": "zeros", "expansion_factor": 0.5, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 5, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=5_expansion=0.5_control_mix=0.0_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=5_expansion=0.5_tokens=24576000_lr=0.001/w0v5k2qm", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=5_expansion=0.5_control=0.1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa4bd21d7419e711c64921d19aff64d00ba5738162d0199d9d6e94ff966a06ad
+size 78015904

l1_coeff=5_expansion=0.5_control=0.1_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": "NeelNanda/openwebtext-tokenized-9b", "is_control_dataset_tokenized": true, "control_mixture": 0.1, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 384, "b_dec_init_method": "zeros", "expansion_factor": 0.5, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 5, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=5_expansion=0.5_control_mix=0.1_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=5_expansion=0.5_tokens=24576000_lr=0.001/7yhypsfb", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=5_expansion=2_control=0.0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ecd7aacff2178848f5166ed1c4f01016b810c66c2563215af2d3f1362a5aa49
+size 85107624

l1_coeff=5_expansion=2_control=0.0_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": null, "is_control_dataset_tokenized": true, "control_mixture": 0.0, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 1536, "b_dec_init_method": "zeros", "expansion_factor": 2, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 5, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=5_expansion=2_control_mix=0.0_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=5_expansion=2_tokens=24576000_lr=0.001/dauks1qd", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=5_expansion=2_control=0.1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:094c750401f20dcf659a2d37d1eb4f23cab146aeb826c04bfdd6c80700b12186
+size 85107624

l1_coeff=5_expansion=2_control=0.1_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": "NeelNanda/openwebtext-tokenized-9b", "is_control_dataset_tokenized": true, "control_mixture": 0.1, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 1536, "b_dec_init_method": "zeros", "expansion_factor": 2, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 5, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=5_expansion=2_control_mix=0.1_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=5_expansion=2_tokens=24576000_lr=0.001/u3tveb52", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=6_expansion=0.5_control=0.0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:167f8c846a4211f2afd4858daadcd76303062ac3848791b55d764e29ac1381aa
+size 78015904

l1_coeff=6_expansion=0.5_control=0.0_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": null, "is_control_dataset_tokenized": true, "control_mixture": 0.0, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 384, "b_dec_init_method": "zeros", "expansion_factor": 0.5, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 6, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=6_expansion=0.5_control_mix=0.0_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=6_expansion=0.5_tokens=24576000_lr=0.001/30ds4e5j", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=6_expansion=0.5_control=0.1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f57df5fc1ab6ad42fc2a946287217535629c07880e3703a39cdb03ea494ae645
+size 78015904

l1_coeff=6_expansion=0.5_control=0.1_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": "NeelNanda/openwebtext-tokenized-9b", "is_control_dataset_tokenized": true, "control_mixture": 0.1, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 384, "b_dec_init_method": "zeros", "expansion_factor": 0.5, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 6, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=6_expansion=0.5_control_mix=0.1_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=6_expansion=0.5_tokens=24576000_lr=0.001/5r41rymj", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=6_expansion=2_control=0.0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13050cc2f067b4c15a15411b835bb73e0357a6d646d8b1b27ff6faf714a3462c
+size 85107624

l1_coeff=6_expansion=2_control=0.0_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": null, "is_control_dataset_tokenized": true, "control_mixture": 0.0, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 1536, "b_dec_init_method": "zeros", "expansion_factor": 2, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 6, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=6_expansion=2_control_mix=0.0_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=6_expansion=2_tokens=24576000_lr=0.001/tx167ps2", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}

l1_coeff=6_expansion=2_control=0.1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81e2928c0bbb3c00b8949a80fb10d37f1bfe9c9db65c1408409684402ff0836a
+size 85107624

l1_coeff=6_expansion=2_control=0.1_cfg.json

@@ -0,0 +1 @@
+{"gsae_repo": "jacobcd52/mats=gpt2-saes", "gsae_filename_no_suffix": "gpt2-small_l1_coeff=45_expansion=64_tokens=409600000_lr=1e-05", "control_dataset_path": "NeelNanda/openwebtext-tokenized-9b", "is_control_dataset_tokenized": true, "control_mixture": 0.1, "model_name": "gpt2-small", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_pre", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "jacobcd52/physics-papers", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 256, "use_cached_activations": false, "cached_activations_path": null, "architecture": "gated", "d_in": 768, "d_sae": 1536, "b_dec_init_method": "zeros", "expansion_factor": 2, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 24576000, "finetuning_tokens": 0, "store_batch_size_prompts": 16, "train_batch_size_tokens": 8192, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 6, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 150, "lr": 0.001, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 0.0001, "lr_decay_steps": 600, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "phys-SSAE-wide-gsae", "wandb_id": null, "run_name": "l1=6_expansion=2_control_mix=0.1_tokens=24576000_lr=0.001", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 20, "resume": false, "n_checkpoints": 0, "checkpoint_path": "gpt2_gsae_l1_coeff=6_expansion=2_tokens=24576000_lr=0.001/qbxbkel8", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.7.0", "sae_lens_training_version": "3.7.0", "tokens_per_buffer": 134217728}