diff --git a/mul_fractal_2_16_l0_128/config.json b/mul_fractal_2_16_l0_128/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..add01f98eb4d162fb4f8f736343b4216ea1c0267
--- /dev/null
+++ b/mul_fractal_2_16_l0_128/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 128,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 2,
+    "num_nkeys": 16,
+    "num_heads": 2048,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 888362139,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 2048, 'num_mkeys': 2, 'num_nkeys': 16, 'num_tokens': 888362139, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_2_16_l0_128/sae.pt b/mul_fractal_2_16_l0_128/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..0b042d027857e74c4d59984e3f90753e02049164
--- /dev/null
+++ b/mul_fractal_2_16_l0_128/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f59b70805f5927fa1b119dee146475f1a78d7afcc9f55f06b98137d935a7aa2a
+size 944139608
diff --git a/mul_fractal_2_16_l0_16/config.json b/mul_fractal_2_16_l0_16/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..01896fba49b9dfd9a0668fbc89ca025dab8528eb
--- /dev/null
+++ b/mul_fractal_2_16_l0_16/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 16,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 2,
+    "num_nkeys": 16,
+    "num_heads": 2048,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 888362139,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 2048, 'num_mkeys': 2, 'num_nkeys': 16, 'num_tokens': 888362139, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_2_16_l0_16/sae.pt b/mul_fractal_2_16_l0_16/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..3bef1ed7974d541df9251fd8e4a9bc79f17d409a
--- /dev/null
+++ b/mul_fractal_2_16_l0_16/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0de8b3f8ccd77047cf6126c254aec2227f9243a40d52e5d1a2078e9b6fbf2bee
+size 944139608
diff --git a/mul_fractal_2_16_l0_256/config.json b/mul_fractal_2_16_l0_256/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..1de2934997d1179290776693a4ac98b75a72cce9
--- /dev/null
+++ b/mul_fractal_2_16_l0_256/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 256,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 2,
+    "num_nkeys": 16,
+    "num_heads": 2048,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 888362139,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 2048, 'num_mkeys': 2, 'num_nkeys': 16, 'num_tokens': 888362139, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_2_16_l0_256/sae.pt b/mul_fractal_2_16_l0_256/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..a46792e1cf01f91ca33be1a6ea2831eaa231e172
--- /dev/null
+++ b/mul_fractal_2_16_l0_256/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8590bfe9f0626751106a6f5a9d84c0066671b5f01c5d7ddfda3079101d792b6
+size 944139608
diff --git a/mul_fractal_2_16_l0_32/config.json b/mul_fractal_2_16_l0_32/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..f89d130d929cd0ef18301ac5d76309d6340702a0
--- /dev/null
+++ b/mul_fractal_2_16_l0_32/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 32,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 2,
+    "num_nkeys": 16,
+    "num_heads": 2048,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 888362139,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 2048, 'num_mkeys': 2, 'num_nkeys': 16, 'num_tokens': 888362139, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_2_16_l0_32/sae.pt b/mul_fractal_2_16_l0_32/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..5471ff65fcc2fbc8023bfc11bde23a658bff5f1b
--- /dev/null
+++ b/mul_fractal_2_16_l0_32/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:68691d277ebc3c3b219f7ca214b44612dbf45b954a0cb61a109abbdd9976541f
+size 944139608
diff --git a/mul_fractal_2_16_l0_64/config.json b/mul_fractal_2_16_l0_64/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..c5288faf4f1b629b45879db37f2a3b073966e1b5
--- /dev/null
+++ b/mul_fractal_2_16_l0_64/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 64,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 2,
+    "num_nkeys": 16,
+    "num_heads": 2048,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 888362139,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 2048, 'num_mkeys': 2, 'num_nkeys': 16, 'num_tokens': 888362139, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_2_16_l0_64/sae.pt b/mul_fractal_2_16_l0_64/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..0d186a5f56f14b1f794825eed0f9c7ccc631ddb9
--- /dev/null
+++ b/mul_fractal_2_16_l0_64/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a3d5cf28d6bd75dffa482733b1a05c88b04ccc004ecee22a98f1aa6b329e025
+size 944139608
diff --git a/mul_fractal_2_4_l0_128/config.json b/mul_fractal_2_4_l0_128/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..a0e08b599c0af1e969e06d76f95b4a6003c66366
--- /dev/null
+++ b/mul_fractal_2_4_l0_128/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 128,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 2,
+    "num_nkeys": 4,
+    "num_heads": 8192,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 666497296,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 8192, 'num_mkeys': 2, 'num_nkeys': 4, 'num_tokens': 666497296, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_2_4_l0_128/sae.pt b/mul_fractal_2_4_l0_128/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..4d6caf74fa9017734bcfe1594ed77c6b5fd90ca9
--- /dev/null
+++ b/mul_fractal_2_4_l0_128/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e28af90a4206ab9011daa741e24aab8afee51668422492ee88653e895e47424
+size 1057434968
diff --git a/mul_fractal_2_4_l0_16/config.json b/mul_fractal_2_4_l0_16/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..a0517ad79ca909e35d7522d82d8466888d929d0a
--- /dev/null
+++ b/mul_fractal_2_4_l0_16/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 16,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 2,
+    "num_nkeys": 4,
+    "num_heads": 8192,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 666497296,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 8192, 'num_mkeys': 2, 'num_nkeys': 4, 'num_tokens': 666497296, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_2_4_l0_16/sae.pt b/mul_fractal_2_4_l0_16/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..8b5e51b2066d74d66064e38174ed74a660756f91
--- /dev/null
+++ b/mul_fractal_2_4_l0_16/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0240a85d49b3d247917cc3584b82b4aab87de36405a0c518b48c5a79a00c5bd8
+size 1057434968
diff --git a/mul_fractal_2_4_l0_256/config.json b/mul_fractal_2_4_l0_256/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..e661b297b7e8cd1ed24859b7b4cde0b3a0eed245
--- /dev/null
+++ b/mul_fractal_2_4_l0_256/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 256,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 2,
+    "num_nkeys": 4,
+    "num_heads": 8192,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 666497296,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 8192, 'num_mkeys': 2, 'num_nkeys': 4, 'num_tokens': 666497296, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_2_4_l0_256/sae.pt b/mul_fractal_2_4_l0_256/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..10cdfe218f4948b54f1d4ced38a142057d040c90
--- /dev/null
+++ b/mul_fractal_2_4_l0_256/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f60ce40dec6ad3dee6dc50c9c007d0aaf96fad98ccb7c98411fb68f1b93ccd57
+size 1057434968
diff --git a/mul_fractal_2_4_l0_32/config.json b/mul_fractal_2_4_l0_32/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..b065bd8a227166d9ba6b15d98aedf3f731e5abd3
--- /dev/null
+++ b/mul_fractal_2_4_l0_32/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 32,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 2,
+    "num_nkeys": 4,
+    "num_heads": 8192,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 666497296,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 8192, 'num_mkeys': 2, 'num_nkeys': 4, 'num_tokens': 666497296, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_2_4_l0_32/sae.pt b/mul_fractal_2_4_l0_32/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..247e061e37da14071818e0dab56d2301660b1243
--- /dev/null
+++ b/mul_fractal_2_4_l0_32/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a54317922936455fcec74693bc1ced860ae23610cc9560096377ae548d07a42
+size 1057434968
diff --git a/mul_fractal_2_4_l0_64/config.json b/mul_fractal_2_4_l0_64/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..4b33ffd0efa003e1b338a97be9b9667de13ac26f
--- /dev/null
+++ b/mul_fractal_2_4_l0_64/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 64,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 2,
+    "num_nkeys": 4,
+    "num_heads": 8192,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 666497296,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 8192, 'num_mkeys': 2, 'num_nkeys': 4, 'num_tokens': 666497296, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_2_4_l0_64/sae.pt b/mul_fractal_2_4_l0_64/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..fad6d8a8693d3177e49b8f80992fbcf2937afaa9
--- /dev/null
+++ b/mul_fractal_2_4_l0_64/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3d7542536f062e7ce966d7e9dd50af164bdc0729ff2093a0f2e0f0039c0c404
+size 1057434968
diff --git a/mul_fractal_2_8_l0_128/config.json b/mul_fractal_2_8_l0_128/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..278d8ec52c4a8df25a47a02ab4fff1aaf66c754a
--- /dev/null
+++ b/mul_fractal_2_8_l0_128/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 128,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 2,
+    "num_nkeys": 8,
+    "num_heads": 4096,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 799634235,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 4096, 'num_mkeys': 2, 'num_nkeys': 8, 'num_tokens': 799634235, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_2_8_l0_128/sae.pt b/mul_fractal_2_8_l0_128/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..474b2619eada4ef0692a5e53756ad38f11d09d73
--- /dev/null
+++ b/mul_fractal_2_8_l0_128/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:814fe46c16ad86196b9056ddb4f186567ffd0ccae687b1cc14cc87380c3b27ba
+size 981904728
diff --git a/mul_fractal_2_8_l0_16/config.json b/mul_fractal_2_8_l0_16/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..43296761df0c703145edce1b4f3a47e1e57d9b3c
--- /dev/null
+++ b/mul_fractal_2_8_l0_16/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 16,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 2,
+    "num_nkeys": 8,
+    "num_heads": 4096,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 799634235,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 4096, 'num_mkeys': 2, 'num_nkeys': 8, 'num_tokens': 799634235, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_2_8_l0_16/sae.pt b/mul_fractal_2_8_l0_16/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..9b5b18e536cf281c4943e18991296321a50f3a01
--- /dev/null
+++ b/mul_fractal_2_8_l0_16/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eebf8b2e065e22f886d315f4dcec1671add85daac3160697a99e1196bb55a3f5
+size 981904728
diff --git a/mul_fractal_2_8_l0_256/config.json b/mul_fractal_2_8_l0_256/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..059718b5da491ac463c04b714ff0cd20afd09b9c
--- /dev/null
+++ b/mul_fractal_2_8_l0_256/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 256,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 2,
+    "num_nkeys": 8,
+    "num_heads": 4096,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 799634235,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 4096, 'num_mkeys': 2, 'num_nkeys': 8, 'num_tokens': 799634235, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_2_8_l0_256/sae.pt b/mul_fractal_2_8_l0_256/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..90f07e446314bd8bcb9a2ad916430353c840e2a5
--- /dev/null
+++ b/mul_fractal_2_8_l0_256/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4747708976c209bd290a5a71e66e4877c3aed6417e988d832bdf7d7c1f7e0030
+size 981904728
diff --git a/mul_fractal_2_8_l0_32/config.json b/mul_fractal_2_8_l0_32/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..b678022161616333bbe53288a84b3df73cbc0882
--- /dev/null
+++ b/mul_fractal_2_8_l0_32/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 32,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 2,
+    "num_nkeys": 8,
+    "num_heads": 4096,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 799634235,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 4096, 'num_mkeys': 2, 'num_nkeys': 8, 'num_tokens': 799634235, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_2_8_l0_32/sae.pt b/mul_fractal_2_8_l0_32/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..5ef365416688a817835608ad3e22672f2c8d9970
--- /dev/null
+++ b/mul_fractal_2_8_l0_32/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79abec4705739699f9e0f9ceeecc0b4bb681ba85043a452a24c664ed23aa152c
+size 981904728
diff --git a/mul_fractal_2_8_l0_64/config.json b/mul_fractal_2_8_l0_64/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..78daf007f5dac46ce31e65b44b522bd05117ba00
--- /dev/null
+++ b/mul_fractal_2_8_l0_64/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 64,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 2,
+    "num_nkeys": 8,
+    "num_heads": 4096,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 799634235,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 4096, 'num_mkeys': 2, 'num_nkeys': 8, 'num_tokens': 799634235, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_4_4_l0_128/config.json b/mul_fractal_4_4_l0_128/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..a250a3cc7bce297f367e2d9e61b7bdcfb1c7a81b
--- /dev/null
+++ b/mul_fractal_4_4_l0_128/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 128,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 4,
+    "num_nkeys": 4,
+    "num_heads": 4096,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 999238222,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 4096, 'num_mkeys': 4, 'num_nkeys': 4, 'num_tokens': 999238222, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_4_4_l0_128/sae.pt b/mul_fractal_4_4_l0_128/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..4a142e41c1b417c3de1e7a18624eb0eaf939a227
--- /dev/null
+++ b/mul_fractal_4_4_l0_128/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9534a6632f9e2ea8e58aa4c782aaefe613bc1d626d4a9d75296c092e0a5c5955
+size 906374488
diff --git a/mul_fractal_4_4_l0_16/config.json b/mul_fractal_4_4_l0_16/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..191d7bc1f0c6eb9274a8d51dbd702567b593ef39
--- /dev/null
+++ b/mul_fractal_4_4_l0_16/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 16,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 4,
+    "num_nkeys": 4,
+    "num_heads": 4096,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 999238222,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 4096, 'num_mkeys': 4, 'num_nkeys': 4, 'num_tokens': 999238222, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_4_4_l0_16/sae.pt b/mul_fractal_4_4_l0_16/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..dbbb548d053b291398d1bda245d79e17f82ecb6c
--- /dev/null
+++ b/mul_fractal_4_4_l0_16/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff5f22a8bd962abb1727b834c129831e6c2c694dbb72e299ec256fc87f856e39
+size 906374488
diff --git a/mul_fractal_4_4_l0_256/config.json b/mul_fractal_4_4_l0_256/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..81c9da8cfd461a4f99e30d8d4ab258b58d8c45fc
--- /dev/null
+++ b/mul_fractal_4_4_l0_256/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 256,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 4,
+    "num_nkeys": 4,
+    "num_heads": 4096,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 999238222,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 4096, 'num_mkeys': 4, 'num_nkeys': 4, 'num_tokens': 999238222, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_4_4_l0_256/sae.pt b/mul_fractal_4_4_l0_256/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..1bac7a2c6f298559219f57031c5110dc1c8a0ee7
--- /dev/null
+++ b/mul_fractal_4_4_l0_256/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:91f12b476bb1285b81f4e6d1e6dbc8d77b49ab316c089fa1b0c372851e5e7554
+size 906374488
diff --git a/mul_fractal_4_4_l0_32/config.json b/mul_fractal_4_4_l0_32/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..8d0543f09ff65764c12ef24674190539838d77bf
--- /dev/null
+++ b/mul_fractal_4_4_l0_32/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 32,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 4,
+    "num_nkeys": 4,
+    "num_heads": 4096,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 999238222,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 4096, 'num_mkeys': 4, 'num_nkeys': 4, 'num_tokens': 999238222, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_4_4_l0_32/sae.pt b/mul_fractal_4_4_l0_32/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..32be619a2574058d57261ed9d470f130df952870
--- /dev/null
+++ b/mul_fractal_4_4_l0_32/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:41b9b59f97cadd527ae34183ddd6b5550d407932a2c0a7a11a3f525cf1f08138
+size 906374488
diff --git a/mul_fractal_4_4_l0_64/config.json b/mul_fractal_4_4_l0_64/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..b345460e4f4681f663bf84be8ffd53f69f983a42
--- /dev/null
+++ b/mul_fractal_4_4_l0_64/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 64,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 4,
+    "num_nkeys": 4,
+    "num_heads": 4096,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 999238222,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 4096, 'num_mkeys': 4, 'num_nkeys': 4, 'num_tokens': 999238222, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_4_4_l0_64/sae.pt b/mul_fractal_4_4_l0_64/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..c99a57591ed82790dd57bee72b3b669cfd894825
--- /dev/null
+++ b/mul_fractal_4_4_l0_64/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffb3ae395d3f8cb51a945a470dd79b81a629a07aa2f2327f273f837ce41803fe
+size 906374488
diff --git a/mul_fractal_4_8_l0_128/config.json b/mul_fractal_4_8_l0_128/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..9b7ff63af751146501fa7cb99330bf672c3b1269
--- /dev/null
+++ b/mul_fractal_4_8_l0_128/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 128,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 4,
+    "num_nkeys": 8,
+    "num_heads": 2048,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 1331641354,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 2048, 'num_mkeys': 4, 'num_nkeys': 8, 'num_tokens': 1331641354, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_4_8_l0_128/sae.pt b/mul_fractal_4_8_l0_128/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..83232462ece74a8fbb1fbc8aaf5ea15991943b03
--- /dev/null
+++ b/mul_fractal_4_8_l0_128/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b01988518bb330d6ad5e68964a644411fee10ac1c3eab15a4bbbea2f8b14e0f
+size 830844248
diff --git a/mul_fractal_4_8_l0_16/config.json b/mul_fractal_4_8_l0_16/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..5284a1716f97fc13402e2cf264dd39c4f9c859c6
--- /dev/null
+++ b/mul_fractal_4_8_l0_16/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 16,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 4,
+    "num_nkeys": 8,
+    "num_heads": 2048,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 1331641354,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 2048, 'num_mkeys': 4, 'num_nkeys': 8, 'num_tokens': 1331641354, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_4_8_l0_16/sae.pt b/mul_fractal_4_8_l0_16/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..ad69adb3f4d974444bec9c61b114069f0cb00c87
--- /dev/null
+++ b/mul_fractal_4_8_l0_16/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:68adc681c80db186acfb3ff52c1a915ae8dc06d2a1b74f4f3ae8a22a89e0693f
+size 830844248
diff --git a/mul_fractal_4_8_l0_256/config.json b/mul_fractal_4_8_l0_256/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..79815502eee10d0132d1dd79f79c9aa62778148e
--- /dev/null
+++ b/mul_fractal_4_8_l0_256/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 256,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 4,
+    "num_nkeys": 8,
+    "num_heads": 2048,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 1331641354,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 2048, 'num_mkeys': 4, 'num_nkeys': 8, 'num_tokens': 1331641354, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_4_8_l0_256/sae.pt b/mul_fractal_4_8_l0_256/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..9900332eb4b6209c8b08c7724dd5424cafd92d8d
--- /dev/null
+++ b/mul_fractal_4_8_l0_256/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54785291ec38e9d7bc69a3a57999281a1305ece8ff1ba5421c3866d7bce04787
+size 830844248
diff --git a/mul_fractal_4_8_l0_32/config.json b/mul_fractal_4_8_l0_32/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..7a3a340ea287226445b17e7bbac99a70e8d36405
--- /dev/null
+++ b/mul_fractal_4_8_l0_32/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 32,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 4,
+    "num_nkeys": 8,
+    "num_heads": 2048,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 1331641354,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 2048, 'num_mkeys': 4, 'num_nkeys': 8, 'num_tokens': 1331641354, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_4_8_l0_32/sae.pt b/mul_fractal_4_8_l0_32/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..497de6d8e2a2788ebda51d85fe43d2817d6f6d37
--- /dev/null
+++ b/mul_fractal_4_8_l0_32/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:15e835f4080dea081fc92c9d8c30e0fa69ea8491f98620da5f957b926ddf1100
+size 830844248
diff --git a/mul_fractal_4_8_l0_64/config.json b/mul_fractal_4_8_l0_64/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..cf2c7aec32d458cdf2724e35d9ae96f6792f2c3d
--- /dev/null
+++ b/mul_fractal_4_8_l0_64/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "mul_fractal_topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 64,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": 4,
+    "num_nkeys": 8,
+    "num_heads": 2048,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 1331641354,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_heads': 2048, 'num_mkeys': 4, 'num_nkeys': 8, 'num_tokens': 1331641354, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/mul_fractal_4_8_l0_64/sae.pt b/mul_fractal_4_8_l0_64/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..bd05f7eed5deec8dc4aef00945f51de1ac96577d
--- /dev/null
+++ b/mul_fractal_4_8_l0_64/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea4015a06e64e303b5d95af31bac64cce4ff66e5d1706ea703b086981282b921
+size 830844248
diff --git a/topk_l0_128/config.json b/topk_l0_128/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..d0116a964a3c61c97eb24c63f4c61e19d523b1b7
--- /dev/null
+++ b/topk_l0_128/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 128,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": null,
+    "num_nkeys": null,
+    "num_heads": -1,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 500000000,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_tokens': 500000000, 'sae_type': 'topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/topk_l0_128/sae.pt b/topk_l0_128/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..a8f6302ff4c8bbe63f46e36eb077dbfb29b9ffa0
--- /dev/null
+++ b/topk_l0_128/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ee20d84fd5613248d250aa5c726ba04d5a3a058ff6ef9dc23ae97e5dea33863
+size 1208495512
diff --git a/topk_l0_16/config.json b/topk_l0_16/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..97fc9014332803eb5c894ae795b45cc755345112
--- /dev/null
+++ b/topk_l0_16/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 16,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": null,
+    "num_nkeys": null,
+    "num_heads": -1,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 500000000,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_tokens': 500000000, 'sae_type': 'topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/topk_l0_16/sae.pt b/topk_l0_16/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..46a1453daa59cf7f910b3b65548bf2f426fe2e07
--- /dev/null
+++ b/topk_l0_16/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6db17284bae31eafe44433888d37856569325d510737d967984a61342a0de1c0
+size 1208495512
diff --git a/topk_l0_256/config.json b/topk_l0_256/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..d712fce0d1c3fabf827c36ee00ff130dd8e16b47
--- /dev/null
+++ b/topk_l0_256/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 256,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": null,
+    "num_nkeys": null,
+    "num_heads": -1,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 500000000,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_tokens': 500000000, 'sae_type': 'topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/topk_l0_256/sae.pt b/topk_l0_256/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..700dbf503396abeb366684127eb7902c19192c4e
--- /dev/null
+++ b/topk_l0_256/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:575be21c798bbb11c7084e7dd43577232e703c87493b8fbc660817efd87cdb03
+size 1208495512
diff --git a/topk_l0_32/config.json b/topk_l0_32/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..a5d099eac152732e76372906de1aeb24120aa054
--- /dev/null
+++ b/topk_l0_32/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 32,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": null,
+    "num_nkeys": null,
+    "num_heads": -1,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 500000000,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_tokens': 500000000, 'sae_type': 'topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/topk_l0_32/sae.pt b/topk_l0_32/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..c2790a7bb076570e7cbced9caa4269ac0fc0080f
--- /dev/null
+++ b/topk_l0_32/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:818327d7fea9071c589af2e1ca0163c89d7be3d9b6a6c680b88fd2faa7a53ec2
+size 1208495512
diff --git a/topk_l0_64/config.json b/topk_l0_64/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..90ba51fc067ef2d58ada5a844b31137f2b31d19e
--- /dev/null
+++ b/topk_l0_64/config.json
@@ -0,0 +1,51 @@
+{
+    "model_name": "google/gemma-2-2b",
+    "layer": 12,
+    "hook_point": "resid_post",
+    "act_size": 2304,
+    "sae_type": "topk",
+    "dict_size": 65536,
+    "aux_penalty": 0.03125,
+    "input_unit_norm": true,
+    "batch_norm_on_queries": false,
+    "affine_batch_norm": false,
+    "linear_heads": 0,
+    "topk2": 64,
+    "topk1": 50,
+    "topk2_warmup_steps_fraction": 0.0,
+    "start_topk2": 50,
+    "topk1_warmup_steps_fraction": 0.0,
+    "start_topk1": 50,
+    "topk2_aux": 512,
+    "cartesian_op": "mul",
+    "router_depth": 2,
+    "router_tree_width": null,
+    "num_mkeys": null,
+    "num_nkeys": null,
+    "num_heads": -1,
+    "n_batches_to_dead": 10,
+    "lr": 0.0008,
+    "bandwidth": 0.001,
+    "l1_coeff": 0.0018,
+    "num_tokens": 500000000,
+    "seq_len": 1024,
+    "model_batch_size": 64,
+    "num_batches_in_buffer": 5,
+    "max_grad_norm": 1.0,
+    "batch_size": 8192,
+    "weight_decay": 0.0,
+    "warmup_fraction": 0.1,
+    "scheduler_type": "cosine_with_min_lr",
+    "device": "cuda",
+    "dtype": "torch.float32",
+    "sae_dtype": "torch.float32",
+    "dataset_path": "HuggingFaceFW/fineweb-edu",
+    "wandb_project": "turbo-llama-lens",
+    "enable_wandb": true,
+    "sae_name": "sae",
+    "seed": 42,
+    "performance_log_steps": 100,
+    "save_checkpoint_steps": 15000000,
+    "wandb_run_suffix": "ex72_for_sae_bench_gemma",
+    "sweep_pair": "{'dict_size': 65536, 'num_tokens': 500000000, 'sae_type': 'topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
+}
\ No newline at end of file
diff --git a/topk_l0_64/sae.pt b/topk_l0_64/sae.pt
new file mode 100644
index 0000000000000000000000000000000000000000..5f3b43d5599c42e81a79ecb783a9ffa0dad8670f
--- /dev/null
+++ b/topk_l0_64/sae.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:85e567aedf6c31471b035d2c738bb12ed99c56ccfbc488f36f74946d914f91d8
+size 1208495512