Add reverse SAEs

hook_embed/reversed/cfg.json

@@ -0,0 +1 @@
+{"model_name": "1l-gelu", "model_class_name": "HookedTransformer", "hook_name": "hook_embed", "hook_eval": "NOT_IN_USE", "hook_layer": 0, "hook_head_index": null, "dataset_path": "ghidav/arithmetics_reversed", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 18, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 512, "d_sae": 8192, "b_dec_init_method": "zeros", "expansion_factor": 16, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 50000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 1024, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 0.1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 2441, "lr": 0.0005, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 5e-05, "lr_decay_steps": 9765, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "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": "sae-feature-circuits", "wandb_id": null, "run_name": "L0_hook_embed_L1_0_1_rev", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 0, "checkpoint_path": "checkpoints/9uer8ihp", "verbose": false, "model_kwargs": {}, "model_from_pretrained_kwargs": {"center_writing_weights": false}, "sae_lens_version": "3.20.5", "sae_lens_training_version": "3.20.5", "tokens_per_buffer": 2359296}

hook_embed/reversed/sae_weights.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cac0f6b9d84f6cfc270357db02224699aeced8ff064c2c96b478c2492bec298f
+size 33622400

hook_embed/reversed/sparsity.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc0fc30655809e2e93371d4a8e02fe4ce4c721e3a748e3b7b03b2c5fdbecc31b
+size 32848

l1_2/reversed/hook_mlp_out/cfg.json

@@ -0,0 +1,86 @@
+{
+    "model_name": "1l-gelu",
+    "model_class_name": "HookedTransformer",
+    "hook_name": "blocks.0.hook_mlp_out",
+    "hook_eval": "NOT_IN_USE",
+    "hook_layer": 0,
+    "hook_head_index": null,
+    "dataset_path": "ghidav/arithmetics_reversed",
+    "dataset_trust_remote_code": true,
+    "streaming": true,
+    "is_dataset_tokenized": true,
+    "context_size": 18,
+    "use_cached_activations": false,
+    "cached_activations_path": null,
+    "architecture": "jumprelu",
+    "d_in": 512,
+    "d_sae": 8192,
+    "b_dec_init_method": "zeros",
+    "expansion_factor": 16,
+    "activation_fn": "relu",
+    "activation_fn_kwargs": {},
+    "normalize_sae_decoder": true,
+    "noise_scale": 0.0,
+    "from_pretrained_path": null,
+    "apply_b_dec_to_input": false,
+    "decoder_orthogonal_init": false,
+    "decoder_heuristic_init": false,
+    "init_encoder_as_decoder_transpose": false,
+    "n_batches_in_buffer": 128,
+    "training_tokens": 50000000,
+    "finetuning_tokens": 0,
+    "store_batch_size_prompts": 8,
+    "train_batch_size_tokens": 1024,
+    "normalize_activations": "none",
+    "device": "cuda",
+    "act_store_device": "cuda",
+    "seed": 42,
+    "dtype": "float32",
+    "prepend_bos": false,
+    "autocast": false,
+    "autocast_lm": false,
+    "compile_llm": false,
+    "llm_compilation_mode": null,
+    "compile_sae": false,
+    "sae_compilation_mode": null,
+    "adam_beta1": 0,
+    "adam_beta2": 0.999,
+    "mse_loss_normalization": null,
+    "l1_coefficient": 2.0,
+    "lp_norm": 1,
+    "scale_sparsity_penalty_by_decoder_norm": false,
+    "l1_warm_up_steps": 2441,
+    "lr": 0.0005,
+    "lr_scheduler_name": "constant",
+    "lr_warm_up_steps": 0,
+    "lr_end": 5e-05,
+    "lr_decay_steps": 9765,
+    "n_restart_cycles": 1,
+    "finetuning_method": null,
+    "use_ghost_grads": false,
+    "feature_sampling_window": 2000,
+    "dead_feature_window": 1000,
+    "dead_feature_threshold": 1e-06,
+    "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": "sae-feature-circuits",
+    "wandb_id": null,
+    "run_name": "L0_hook_mlp_out_L1_2_0_rev",
+    "wandb_entity": null,
+    "wandb_log_frequency": 30,
+    "eval_every_n_wandb_logs": 100,
+    "resume": false,
+    "n_checkpoints": 0,
+    "checkpoint_path": "checkpoints/fhisi0p1",
+    "verbose": false,
+    "model_kwargs": {},
+    "model_from_pretrained_kwargs": {
+        "center_writing_weights": false
+    },
+    "sae_lens_version": "3.20.5",
+    "sae_lens_training_version": "3.20.5",
+    "tokens_per_buffer": 2359296
+}

l1_2/reversed/hook_mlp_out/sae_weights.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a72f511cee7d7bd45340e360ba305dbf01acc7e4152f15d5de8ff7342e31fd78
+size 33622400

l1_2/reversed/hook_mlp_out/sparsity.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:847b1f885ed410c120619b963920e805d45f77dc641c0a2908c52ba1a0eb5df6
+size 32848

l1_2/reversed/hook_resid_post/cfg.json

@@ -0,0 +1,86 @@
+{
+    "model_name": "1l-gelu",
+    "model_class_name": "HookedTransformer",
+    "hook_name": "blocks.0.hook_resid_post",
+    "hook_eval": "NOT_IN_USE",
+    "hook_layer": 0,
+    "hook_head_index": null,
+    "dataset_path": "ghidav/arithmetics_reversed",
+    "dataset_trust_remote_code": true,
+    "streaming": true,
+    "is_dataset_tokenized": true,
+    "context_size": 18,
+    "use_cached_activations": false,
+    "cached_activations_path": null,
+    "architecture": "jumprelu",
+    "d_in": 512,
+    "d_sae": 8192,
+    "b_dec_init_method": "zeros",
+    "expansion_factor": 16,
+    "activation_fn": "relu",
+    "activation_fn_kwargs": {},
+    "normalize_sae_decoder": true,
+    "noise_scale": 0.0,
+    "from_pretrained_path": null,
+    "apply_b_dec_to_input": false,
+    "decoder_orthogonal_init": false,
+    "decoder_heuristic_init": false,
+    "init_encoder_as_decoder_transpose": false,
+    "n_batches_in_buffer": 128,
+    "training_tokens": 50000000,
+    "finetuning_tokens": 0,
+    "store_batch_size_prompts": 8,
+    "train_batch_size_tokens": 1024,
+    "normalize_activations": "none",
+    "device": "cuda",
+    "act_store_device": "cuda",
+    "seed": 42,
+    "dtype": "float32",
+    "prepend_bos": false,
+    "autocast": false,
+    "autocast_lm": false,
+    "compile_llm": false,
+    "llm_compilation_mode": null,
+    "compile_sae": false,
+    "sae_compilation_mode": null,
+    "adam_beta1": 0,
+    "adam_beta2": 0.999,
+    "mse_loss_normalization": null,
+    "l1_coefficient": 2.0,
+    "lp_norm": 1,
+    "scale_sparsity_penalty_by_decoder_norm": false,
+    "l1_warm_up_steps": 2441,
+    "lr": 0.0005,
+    "lr_scheduler_name": "constant",
+    "lr_warm_up_steps": 0,
+    "lr_end": 5e-05,
+    "lr_decay_steps": 9765,
+    "n_restart_cycles": 1,
+    "finetuning_method": null,
+    "use_ghost_grads": false,
+    "feature_sampling_window": 2000,
+    "dead_feature_window": 1000,
+    "dead_feature_threshold": 1e-06,
+    "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": "sae-feature-circuits",
+    "wandb_id": null,
+    "run_name": "L0_hook_resid_post_L1_2_0_rev",
+    "wandb_entity": null,
+    "wandb_log_frequency": 30,
+    "eval_every_n_wandb_logs": 100,
+    "resume": false,
+    "n_checkpoints": 0,
+    "checkpoint_path": "checkpoints/220duyfn",
+    "verbose": false,
+    "model_kwargs": {},
+    "model_from_pretrained_kwargs": {
+        "center_writing_weights": false
+    },
+    "sae_lens_version": "3.20.5",
+    "sae_lens_training_version": "3.20.5",
+    "tokens_per_buffer": 2359296
+}

l1_2/reversed/hook_resid_post/sae_weights.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e9591b30b7fad6cc025b6fc149a7f24454336ec4fcf0b82abb0a0a4573c06cc2
+size 33622400

l1_2/reversed/hook_resid_post/sparsity.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44d9853721551165bdc0fce8fff74366cb63c078c2ad668dc8780bef79dcf64c
+size 32848

l1_2/reversed/hook_z/cfg.json

@@ -0,0 +1,86 @@
+{
+    "model_name": "1l-gelu",
+    "model_class_name": "HookedTransformer",
+    "hook_name": "blocks.0.attn.hook_z",
+    "hook_eval": "NOT_IN_USE",
+    "hook_layer": 0,
+    "hook_head_index": null,
+    "dataset_path": "ghidav/arithmetics_reversed",
+    "dataset_trust_remote_code": true,
+    "streaming": true,
+    "is_dataset_tokenized": true,
+    "context_size": 18,
+    "use_cached_activations": false,
+    "cached_activations_path": null,
+    "architecture": "jumprelu",
+    "d_in": 512,
+    "d_sae": 8192,
+    "b_dec_init_method": "zeros",
+    "expansion_factor": 16,
+    "activation_fn": "relu",
+    "activation_fn_kwargs": {},
+    "normalize_sae_decoder": true,
+    "noise_scale": 0.0,
+    "from_pretrained_path": null,
+    "apply_b_dec_to_input": false,
+    "decoder_orthogonal_init": false,
+    "decoder_heuristic_init": false,
+    "init_encoder_as_decoder_transpose": false,
+    "n_batches_in_buffer": 128,
+    "training_tokens": 50000000,
+    "finetuning_tokens": 0,
+    "store_batch_size_prompts": 8,
+    "train_batch_size_tokens": 1024,
+    "normalize_activations": "none",
+    "device": "cuda",
+    "act_store_device": "cuda",
+    "seed": 42,
+    "dtype": "float32",
+    "prepend_bos": false,
+    "autocast": false,
+    "autocast_lm": false,
+    "compile_llm": false,
+    "llm_compilation_mode": null,
+    "compile_sae": false,
+    "sae_compilation_mode": null,
+    "adam_beta1": 0,
+    "adam_beta2": 0.999,
+    "mse_loss_normalization": null,
+    "l1_coefficient": 2.0,
+    "lp_norm": 1,
+    "scale_sparsity_penalty_by_decoder_norm": false,
+    "l1_warm_up_steps": 2441,
+    "lr": 0.0005,
+    "lr_scheduler_name": "constant",
+    "lr_warm_up_steps": 0,
+    "lr_end": 5e-05,
+    "lr_decay_steps": 9765,
+    "n_restart_cycles": 1,
+    "finetuning_method": null,
+    "use_ghost_grads": false,
+    "feature_sampling_window": 2000,
+    "dead_feature_window": 1000,
+    "dead_feature_threshold": 1e-06,
+    "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": "sae-feature-circuits",
+    "wandb_id": null,
+    "run_name": "L0_attn.hook_z_L1_2_0_rev",
+    "wandb_entity": null,
+    "wandb_log_frequency": 30,
+    "eval_every_n_wandb_logs": 100,
+    "resume": false,
+    "n_checkpoints": 0,
+    "checkpoint_path": "checkpoints/lb56kz30",
+    "verbose": false,
+    "model_kwargs": {},
+    "model_from_pretrained_kwargs": {
+        "center_writing_weights": false
+    },
+    "sae_lens_version": "3.20.5",
+    "sae_lens_training_version": "3.20.5",
+    "tokens_per_buffer": 2359296
+}

l1_2/reversed/hook_z/sae_weights.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a43785585f70699841ba234476faa025b08732b43bde62317d5b882af8be8282
+size 33622400

l1_2/reversed/hook_z/sparsity.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e961506e87446115d1621831806ee8deed48e19432982b5ec90ed22e1dacb203
+size 32848