Upload folder using huggingface_hub

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_0/ae.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d7a7b5ebe8b219620bf8c1c652d2b4b42418b66b986b67a583a2840f216bee8
+size 1208233552

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_0/config.json

@@ -0,0 +1,53 @@
+{
+    "trainer": {
+        "trainer_class": "MatryoshkaBatchTopKTrainer",
+        "dict_class": "MatryoshkaBatchTopKSAE",
+        "lr": 0.0002,
+        "steps": 244140,
+        "auxk_alpha": 0.03125,
+        "warmup_steps": 1000,
+        "decay_start": 195312,
+        "threshold_beta": 0.999,
+        "threshold_start_step": 1000,
+        "top_k_aux": 1152,
+        "seed": 1,
+        "activation_dim": 2304,
+        "dict_size": 65536,
+        "group_fractions": [
+            0.03125,
+            0.0625,
+            0.125,
+            0.25,
+            0.53125
+        ],
+        "group_weights": [
+            0.2,
+            0.2,
+            0.2,
+            0.2,
+            0.2
+        ],
+        "group_sizes": [
+            2048,
+            4096,
+            8192,
+            16384,
+            34816
+        ],
+        "k": 25,
+        "device": "cuda:0",
+        "layer": 12,
+        "lm_name": "google/gemma-2-2b",
+        "wandb_name": "MatryoshkaBatchTopKTrainer-google/gemma-2-2b-resid_post_layer_12_trainer_0",
+        "submodule_name": "resid_post_layer_12"
+    },
+    "buffer": {
+        "d_submodule": 2304,
+        "io": "out",
+        "n_ctxs": 244,
+        "ctx_len": 1024,
+        "refresh_batch_size": 48,
+        "out_batch_size": 2048,
+        "device": "cuda:0"
+    }
+}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_0/eval_results.json

@@ -0,0 +1 @@
+{"l2_loss": 61.71621621621622, "l1_loss": 343.4594594594595, "l0": 25.02615920058242, "frac_variance_explained": 0.7965230855855856, "cossim": 0.90234375, "l2_ratio": 0.9157164977477478, "relative_reconstruction_bias": 1.0092201576576576, "loss_original": 2.319045608108108, "loss_reconstructed": 2.5043637387387387, "loss_zero": 12.4375, "frac_recovered": 0.9818060247747747, "frac_alive": 0.907073974609375, "hyperparameters": {"n_inputs": 1000, "context_length": 1024}}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_1/ae.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83545308e48a5252ace9507d0b9a8c163397bc9da4bb7e77eab04cde16375018
+size 1208233552

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_1/config.json

@@ -0,0 +1,53 @@
+{
+    "trainer": {
+        "trainer_class": "MatryoshkaBatchTopKTrainer",
+        "dict_class": "MatryoshkaBatchTopKSAE",
+        "lr": 0.0002,
+        "steps": 244140,
+        "auxk_alpha": 0.03125,
+        "warmup_steps": 1000,
+        "decay_start": 195312,
+        "threshold_beta": 0.999,
+        "threshold_start_step": 1000,
+        "top_k_aux": 1152,
+        "seed": 1,
+        "activation_dim": 2304,
+        "dict_size": 65536,
+        "group_fractions": [
+            0.03125,
+            0.0625,
+            0.125,
+            0.25,
+            0.53125
+        ],
+        "group_weights": [
+            0.2,
+            0.2,
+            0.2,
+            0.2,
+            0.2
+        ],
+        "group_sizes": [
+            2048,
+            4096,
+            8192,
+            16384,
+            34816
+        ],
+        "k": 40,
+        "device": "cuda:0",
+        "layer": 12,
+        "lm_name": "google/gemma-2-2b",
+        "wandb_name": "MatryoshkaBatchTopKTrainer-google/gemma-2-2b-resid_post_layer_12_trainer_1",
+        "submodule_name": "resid_post_layer_12"
+    },
+    "buffer": {
+        "d_submodule": 2304,
+        "io": "out",
+        "n_ctxs": 244,
+        "ctx_len": 1024,
+        "refresh_batch_size": 48,
+        "out_batch_size": 2048,
+        "device": "cuda:0"
+    }
+}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_1/eval_results.json

@@ -0,0 +1 @@
+{"l2_loss": 57.8490990990991, "l1_loss": 424.1261261261261, "l0": 39.67530561567427, "frac_variance_explained": 0.8202069256756757, "cossim": 0.9140625, "l2_ratio": 0.9258516328828829, "relative_reconstruction_bias": 1.0080236486486487, "loss_original": 2.319045608108108, "loss_reconstructed": 2.4353885135135136, "loss_zero": 12.4375, "frac_recovered": 0.9888443130630631, "frac_alive": 0.9098663330078125, "hyperparameters": {"n_inputs": 1000, "context_length": 1024}}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_10/config.json

@@ -0,0 +1,53 @@
+{
+    "trainer": {
+        "trainer_class": "MatryoshkaBatchTopKTrainer",
+        "dict_class": "MatryoshkaBatchTopKSAE",
+        "lr": 0.0002,
+        "steps": 244140,
+        "auxk_alpha": 0.03125,
+        "warmup_steps": 1000,
+        "decay_start": 195312,
+        "threshold_beta": 0.999,
+        "threshold_start_step": 1000,
+        "top_k_aux": 1152,
+        "seed": 721077,
+        "activation_dim": 2304,
+        "dict_size": 65536,
+        "group_fractions": [
+            0.03125,
+            0.0625,
+            0.125,
+            0.25,
+            0.53125
+        ],
+        "group_weights": [
+            0.2,
+            0.2,
+            0.2,
+            0.2,
+            0.2
+        ],
+        "group_sizes": [
+            2048,
+            4096,
+            8192,
+            16384,
+            34816
+        ],
+        "k": 55,
+        "device": "cuda:0",
+        "layer": 12,
+        "lm_name": "google/gemma-2-2b",
+        "wandb_name": "MatryoshkaBatchTopKTrainer-google/gemma-2-2b-resid_post_layer_12_trainer_10",
+        "submodule_name": "resid_post_layer_12"
+    },
+    "buffer": {
+        "d_submodule": 2304,
+        "io": "out",
+        "n_ctxs": 244,
+        "ctx_len": 1024,
+        "refresh_batch_size": 32,
+        "out_batch_size": 2048,
+        "device": "cuda:0"
+    }
+}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_11/config.json

@@ -0,0 +1,53 @@
+{
+    "trainer": {
+        "trainer_class": "MatryoshkaBatchTopKTrainer",
+        "dict_class": "MatryoshkaBatchTopKSAE",
+        "lr": 0.0002,
+        "steps": 244140,
+        "auxk_alpha": 0.03125,
+        "warmup_steps": 1000,
+        "decay_start": 195312,
+        "threshold_beta": 0.999,
+        "threshold_start_step": 1000,
+        "top_k_aux": 1152,
+        "seed": 721077,
+        "activation_dim": 2304,
+        "dict_size": 65536,
+        "group_fractions": [
+            0.03125,
+            0.0625,
+            0.125,
+            0.25,
+            0.53125
+        ],
+        "group_weights": [
+            0.2,
+            0.2,
+            0.2,
+            0.2,
+            0.2
+        ],
+        "group_sizes": [
+            2048,
+            4096,
+            8192,
+            16384,
+            34816
+        ],
+        "k": 70,
+        "device": "cuda:0",
+        "layer": 12,
+        "lm_name": "google/gemma-2-2b",
+        "wandb_name": "MatryoshkaBatchTopKTrainer-google/gemma-2-2b-resid_post_layer_12_trainer_11",
+        "submodule_name": "resid_post_layer_12"
+    },
+    "buffer": {
+        "d_submodule": 2304,
+        "io": "out",
+        "n_ctxs": 244,
+        "ctx_len": 1024,
+        "refresh_batch_size": 32,
+        "out_batch_size": 2048,
+        "device": "cuda:0"
+    }
+}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_12/config.json

@@ -0,0 +1,53 @@
+{
+    "trainer": {
+        "trainer_class": "MatryoshkaBatchTopKTrainer",
+        "dict_class": "MatryoshkaBatchTopKSAE",
+        "lr": 0.0002,
+        "steps": 244140,
+        "auxk_alpha": 0.03125,
+        "warmup_steps": 1000,
+        "decay_start": 195312,
+        "threshold_beta": 0.999,
+        "threshold_start_step": 1000,
+        "top_k_aux": 1152,
+        "seed": 721077,
+        "activation_dim": 2304,
+        "dict_size": 65536,
+        "group_fractions": [
+            0.03125,
+            0.0625,
+            0.125,
+            0.25,
+            0.53125
+        ],
+        "group_weights": [
+            0.2,
+            0.2,
+            0.2,
+            0.2,
+            0.2
+        ],
+        "group_sizes": [
+            2048,
+            4096,
+            8192,
+            16384,
+            34816
+        ],
+        "k": 85,
+        "device": "cuda:0",
+        "layer": 12,
+        "lm_name": "google/gemma-2-2b",
+        "wandb_name": "MatryoshkaBatchTopKTrainer-google/gemma-2-2b-resid_post_layer_12_trainer_12",
+        "submodule_name": "resid_post_layer_12"
+    },
+    "buffer": {
+        "d_submodule": 2304,
+        "io": "out",
+        "n_ctxs": 244,
+        "ctx_len": 1024,
+        "refresh_batch_size": 32,
+        "out_batch_size": 2048,
+        "device": "cuda:0"
+    }
+}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_13/config.json

@@ -0,0 +1,53 @@
+{
+    "trainer": {
+        "trainer_class": "MatryoshkaBatchTopKTrainer",
+        "dict_class": "MatryoshkaBatchTopKSAE",
+        "lr": 0.0002,
+        "steps": 244140,
+        "auxk_alpha": 0.03125,
+        "warmup_steps": 1000,
+        "decay_start": 195312,
+        "threshold_beta": 0.999,
+        "threshold_start_step": 1000,
+        "top_k_aux": 1152,
+        "seed": 721077,
+        "activation_dim": 2304,
+        "dict_size": 65536,
+        "group_fractions": [
+            0.03125,
+            0.0625,
+            0.125,
+            0.25,
+            0.53125
+        ],
+        "group_weights": [
+            0.2,
+            0.2,
+            0.2,
+            0.2,
+            0.2
+        ],
+        "group_sizes": [
+            2048,
+            4096,
+            8192,
+            16384,
+            34816
+        ],
+        "k": 100,
+        "device": "cuda:0",
+        "layer": 12,
+        "lm_name": "google/gemma-2-2b",
+        "wandb_name": "MatryoshkaBatchTopKTrainer-google/gemma-2-2b-resid_post_layer_12_trainer_13",
+        "submodule_name": "resid_post_layer_12"
+    },
+    "buffer": {
+        "d_submodule": 2304,
+        "io": "out",
+        "n_ctxs": 244,
+        "ctx_len": 1024,
+        "refresh_batch_size": 32,
+        "out_batch_size": 2048,
+        "device": "cuda:0"
+    }
+}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_14/config.json

@@ -0,0 +1,53 @@
+{
+    "trainer": {
+        "trainer_class": "MatryoshkaBatchTopKTrainer",
+        "dict_class": "MatryoshkaBatchTopKSAE",
+        "lr": 0.0002,
+        "steps": 244140,
+        "auxk_alpha": 0.03125,
+        "warmup_steps": 1000,
+        "decay_start": 195312,
+        "threshold_beta": 0.999,
+        "threshold_start_step": 1000,
+        "top_k_aux": 1152,
+        "seed": 721077,
+        "activation_dim": 2304,
+        "dict_size": 65536,
+        "group_fractions": [
+            0.03125,
+            0.0625,
+            0.125,
+            0.25,
+            0.53125
+        ],
+        "group_weights": [
+            0.2,
+            0.2,
+            0.2,
+            0.2,
+            0.2
+        ],
+        "group_sizes": [
+            2048,
+            4096,
+            8192,
+            16384,
+            34816
+        ],
+        "k": 115,
+        "device": "cuda:0",
+        "layer": 12,
+        "lm_name": "google/gemma-2-2b",
+        "wandb_name": "MatryoshkaBatchTopKTrainer-google/gemma-2-2b-resid_post_layer_12_trainer_14",
+        "submodule_name": "resid_post_layer_12"
+    },
+    "buffer": {
+        "d_submodule": 2304,
+        "io": "out",
+        "n_ctxs": 244,
+        "ctx_len": 1024,
+        "refresh_batch_size": 32,
+        "out_batch_size": 2048,
+        "device": "cuda:0"
+    }
+}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_15/config.json

@@ -0,0 +1,53 @@
+{
+    "trainer": {
+        "trainer_class": "MatryoshkaBatchTopKTrainer",
+        "dict_class": "MatryoshkaBatchTopKSAE",
+        "lr": 0.0002,
+        "steps": 244140,
+        "auxk_alpha": 0.03125,
+        "warmup_steps": 1000,
+        "decay_start": 195312,
+        "threshold_beta": 0.999,
+        "threshold_start_step": 1000,
+        "top_k_aux": 1152,
+        "seed": 721077,
+        "activation_dim": 2304,
+        "dict_size": 65536,
+        "group_fractions": [
+            0.03125,
+            0.0625,
+            0.125,
+            0.25,
+            0.53125
+        ],
+        "group_weights": [
+            0.2,
+            0.2,
+            0.2,
+            0.2,
+            0.2
+        ],
+        "group_sizes": [
+            2048,
+            4096,
+            8192,
+            16384,
+            34816
+        ],
+        "k": 130,
+        "device": "cuda:0",
+        "layer": 12,
+        "lm_name": "google/gemma-2-2b",
+        "wandb_name": "MatryoshkaBatchTopKTrainer-google/gemma-2-2b-resid_post_layer_12_trainer_15",
+        "submodule_name": "resid_post_layer_12"
+    },
+    "buffer": {
+        "d_submodule": 2304,
+        "io": "out",
+        "n_ctxs": 244,
+        "ctx_len": 1024,
+        "refresh_batch_size": 32,
+        "out_batch_size": 2048,
+        "device": "cuda:0"
+    }
+}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_2/ae.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d00ec893a3f7cb5cde347f08ca909eb007c4f58266241362fee8fcae5969172e
+size 1208233552

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_2/config.json

@@ -0,0 +1,53 @@
+{
+    "trainer": {
+        "trainer_class": "MatryoshkaBatchTopKTrainer",
+        "dict_class": "MatryoshkaBatchTopKSAE",
+        "lr": 0.0002,
+        "steps": 244140,
+        "auxk_alpha": 0.03125,
+        "warmup_steps": 1000,
+        "decay_start": 195312,
+        "threshold_beta": 0.999,
+        "threshold_start_step": 1000,
+        "top_k_aux": 1152,
+        "seed": 1,
+        "activation_dim": 2304,
+        "dict_size": 65536,
+        "group_fractions": [
+            0.03125,
+            0.0625,
+            0.125,
+            0.25,
+            0.53125
+        ],
+        "group_weights": [
+            0.2,
+            0.2,
+            0.2,
+            0.2,
+            0.2
+        ],
+        "group_sizes": [
+            2048,
+            4096,
+            8192,
+            16384,
+            34816
+        ],
+        "k": 55,
+        "device": "cuda:0",
+        "layer": 12,
+        "lm_name": "google/gemma-2-2b",
+        "wandb_name": "MatryoshkaBatchTopKTrainer-google/gemma-2-2b-resid_post_layer_12_trainer_2",
+        "submodule_name": "resid_post_layer_12"
+    },
+    "buffer": {
+        "d_submodule": 2304,
+        "io": "out",
+        "n_ctxs": 244,
+        "ctx_len": 1024,
+        "refresh_batch_size": 48,
+        "out_batch_size": 2048,
+        "device": "cuda:0"
+    }
+}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_2/eval_results.json

@@ -0,0 +1 @@
+{"l2_loss": 55.37837837837838, "l1_loss": 490.8288288288288, "l0": 54.320920033497856, "frac_variance_explained": 0.8346706081081081, "cossim": 0.921875, "l2_ratio": 0.93359375, "relative_reconstruction_bias": 1.007882882882883, "loss_original": 2.319045608108108, "loss_reconstructed": 2.426097972972973, "loss_zero": 12.4375, "frac_recovered": 0.9893369932432432, "frac_alive": 0.9249725341796875, "hyperparameters": {"n_inputs": 1000, "context_length": 1024}}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_3/ae.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:490e2cf6158c80cbe1615134c0191e3908488c6b35c9205ab9187658917bf004
+size 1208233552

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_3/config.json

@@ -0,0 +1,53 @@
+{
+    "trainer": {
+        "trainer_class": "MatryoshkaBatchTopKTrainer",
+        "dict_class": "MatryoshkaBatchTopKSAE",
+        "lr": 0.0002,
+        "steps": 244140,
+        "auxk_alpha": 0.03125,
+        "warmup_steps": 1000,
+        "decay_start": 195312,
+        "threshold_beta": 0.999,
+        "threshold_start_step": 1000,
+        "top_k_aux": 1152,
+        "seed": 1,
+        "activation_dim": 2304,
+        "dict_size": 65536,
+        "group_fractions": [
+            0.03125,
+            0.0625,
+            0.125,
+            0.25,
+            0.53125
+        ],
+        "group_weights": [
+            0.2,
+            0.2,
+            0.2,
+            0.2,
+            0.2
+        ],
+        "group_sizes": [
+            2048,
+            4096,
+            8192,
+            16384,
+            34816
+        ],
+        "k": 70,
+        "device": "cuda:0",
+        "layer": 12,
+        "lm_name": "google/gemma-2-2b",
+        "wandb_name": "MatryoshkaBatchTopKTrainer-google/gemma-2-2b-resid_post_layer_12_trainer_3",
+        "submodule_name": "resid_post_layer_12"
+    },
+    "buffer": {
+        "d_submodule": 2304,
+        "io": "out",
+        "n_ctxs": 244,
+        "ctx_len": 1024,
+        "refresh_batch_size": 48,
+        "out_batch_size": 2048,
+        "device": "cuda:0"
+    }
+}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_3/eval_results.json

@@ -0,0 +1 @@
+{"l2_loss": 53.48198198198198, "l1_loss": 554.5945945945946, "l0": 69.70713270032728, "frac_variance_explained": 0.8462837837837838, "cossim": 0.9296171171171171, "l2_ratio": 0.94140625, "relative_reconstruction_bias": 1.0092905405405406, "loss_original": 2.319045608108108, "loss_reconstructed": 2.406390765765766, "loss_zero": 12.4375, "frac_recovered": 0.9905686936936937, "frac_alive": 0.9304656982421875, "hyperparameters": {"n_inputs": 1000, "context_length": 1024}}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_4/ae.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2775b748c84e20a05b2e1cfa99cb2067ed476875c88b844e2cc229ea645a24dd
+size 1208233552

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_4/config.json

@@ -0,0 +1,53 @@
+{
+    "trainer": {
+        "trainer_class": "MatryoshkaBatchTopKTrainer",
+        "dict_class": "MatryoshkaBatchTopKSAE",
+        "lr": 0.0002,
+        "steps": 244140,
+        "auxk_alpha": 0.03125,
+        "warmup_steps": 1000,
+        "decay_start": 195312,
+        "threshold_beta": 0.999,
+        "threshold_start_step": 1000,
+        "top_k_aux": 1152,
+        "seed": 1,
+        "activation_dim": 2304,
+        "dict_size": 65536,
+        "group_fractions": [
+            0.03125,
+            0.0625,
+            0.125,
+            0.25,
+            0.53125
+        ],
+        "group_weights": [
+            0.2,
+            0.2,
+            0.2,
+            0.2,
+            0.2
+        ],
+        "group_sizes": [
+            2048,
+            4096,
+            8192,
+            16384,
+            34816
+        ],
+        "k": 85,
+        "device": "cuda:0",
+        "layer": 12,
+        "lm_name": "google/gemma-2-2b",
+        "wandb_name": "MatryoshkaBatchTopKTrainer-google/gemma-2-2b-resid_post_layer_12_trainer_4",
+        "submodule_name": "resid_post_layer_12"
+    },
+    "buffer": {
+        "d_submodule": 2304,
+        "io": "out",
+        "n_ctxs": 244,
+        "ctx_len": 1024,
+        "refresh_batch_size": 48,
+        "out_batch_size": 2048,
+        "device": "cuda:0"
+    }
+}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_4/eval_results.json

@@ -0,0 +1 @@
+{"l2_loss": 52.027027027027025, "l1_loss": 610.8468468468468, "l0": 84.440763112661, "frac_variance_explained": 0.8550816441441441, "cossim": 0.93359375, "l2_ratio": 0.9453125, "relative_reconstruction_bias": 1.0087978603603605, "loss_original": 2.319045608108108, "loss_reconstructed": 2.391328828828829, "loss_zero": 12.4375, "frac_recovered": 0.9915892454954955, "frac_alive": 0.930450439453125, "hyperparameters": {"n_inputs": 1000, "context_length": 1024}}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_5/ae.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b65a0fb98512ee77dc523d0c30062390ed23cb74cc2765bb4c6afca334b4c6f
+size 1208233552

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_5/config.json

@@ -0,0 +1,53 @@
+{
+    "trainer": {
+        "trainer_class": "MatryoshkaBatchTopKTrainer",
+        "dict_class": "MatryoshkaBatchTopKSAE",
+        "lr": 0.0002,
+        "steps": 244140,
+        "auxk_alpha": 0.03125,
+        "warmup_steps": 1000,
+        "decay_start": 195312,
+        "threshold_beta": 0.999,
+        "threshold_start_step": 1000,
+        "top_k_aux": 1152,
+        "seed": 1,
+        "activation_dim": 2304,
+        "dict_size": 65536,
+        "group_fractions": [
+            0.03125,
+            0.0625,
+            0.125,
+            0.25,
+            0.53125
+        ],
+        "group_weights": [
+            0.2,
+            0.2,
+            0.2,
+            0.2,
+            0.2
+        ],
+        "group_sizes": [
+            2048,
+            4096,
+            8192,
+            16384,
+            34816
+        ],
+        "k": 100,
+        "device": "cuda:0",
+        "layer": 12,
+        "lm_name": "google/gemma-2-2b",
+        "wandb_name": "MatryoshkaBatchTopKTrainer-google/gemma-2-2b-resid_post_layer_12_trainer_5",
+        "submodule_name": "resid_post_layer_12"
+    },
+    "buffer": {
+        "d_submodule": 2304,
+        "io": "out",
+        "n_ctxs": 244,
+        "ctx_len": 1024,
+        "refresh_batch_size": 48,
+        "out_batch_size": 2048,
+        "device": "cuda:0"
+    }
+}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_5/eval_results.json

@@ -0,0 +1 @@
+{"l2_loss": 50.80855855855856, "l1_loss": 732.8288288288288, "l0": 99.91276893959389, "frac_variance_explained": 0.8616976351351351, "cossim": 0.9338048986486487, "l2_ratio": 0.94921875, "relative_reconstruction_bias": 1.010135135135135, "loss_original": 2.319045608108108, "loss_reconstructed": 2.3896396396396398, "loss_zero": 12.4375, "frac_recovered": 0.9919411599099099, "frac_alive": 0.93402099609375, "hyperparameters": {"n_inputs": 1000, "context_length": 1024}}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_6/ae.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06bfd16696aa634d2f775f1d280b25ddbfa3452b1d96e48d416971d74f382c74
+size 1208233552

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_6/config.json

@@ -0,0 +1,53 @@
+{
+    "trainer": {
+        "trainer_class": "MatryoshkaBatchTopKTrainer",
+        "dict_class": "MatryoshkaBatchTopKSAE",
+        "lr": 0.0002,
+        "steps": 244140,
+        "auxk_alpha": 0.03125,
+        "warmup_steps": 1000,
+        "decay_start": 195312,
+        "threshold_beta": 0.999,
+        "threshold_start_step": 1000,
+        "top_k_aux": 1152,
+        "seed": 1,
+        "activation_dim": 2304,
+        "dict_size": 65536,
+        "group_fractions": [
+            0.03125,
+            0.0625,
+            0.125,
+            0.25,
+            0.53125
+        ],
+        "group_weights": [
+            0.2,
+            0.2,
+            0.2,
+            0.2,
+            0.2
+        ],
+        "group_sizes": [
+            2048,
+            4096,
+            8192,
+            16384,
+            34816
+        ],
+        "k": 115,
+        "device": "cuda:0",
+        "layer": 12,
+        "lm_name": "google/gemma-2-2b",
+        "wandb_name": "MatryoshkaBatchTopKTrainer-google/gemma-2-2b-resid_post_layer_12_trainer_6",
+        "submodule_name": "resid_post_layer_12"
+    },
+    "buffer": {
+        "d_submodule": 2304,
+        "io": "out",
+        "n_ctxs": 244,
+        "ctx_len": 1024,
+        "refresh_batch_size": 48,
+        "out_batch_size": 2048,
+        "device": "cuda:0"
+    }
+}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_6/eval_results.json

@@ -0,0 +1 @@
+{"l2_loss": 49.817567567567565, "l1_loss": 718.6666666666666, "l0": 114.20686182245478, "frac_variance_explained": 0.8679265202702703, "cossim": 0.9375, "l2_ratio": 0.94921875, "relative_reconstruction_bias": 1.0078125, "loss_original": 2.319045608108108, "loss_reconstructed": 2.369158220720721, "loss_zero": 12.4375, "frac_recovered": 0.9944397522522522, "frac_alive": 0.9305419921875, "hyperparameters": {"n_inputs": 1000, "context_length": 1024}}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_7/ae.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81585fe2d072873fd96a340b7b3434ee089a033c4045f25943e37b8a227e2102
+size 1208233552

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_7/config.json

@@ -0,0 +1,53 @@
+{
+    "trainer": {
+        "trainer_class": "MatryoshkaBatchTopKTrainer",
+        "dict_class": "MatryoshkaBatchTopKSAE",
+        "lr": 0.0002,
+        "steps": 244140,
+        "auxk_alpha": 0.03125,
+        "warmup_steps": 1000,
+        "decay_start": 195312,
+        "threshold_beta": 0.999,
+        "threshold_start_step": 1000,
+        "top_k_aux": 1152,
+        "seed": 1,
+        "activation_dim": 2304,
+        "dict_size": 65536,
+        "group_fractions": [
+            0.03125,
+            0.0625,
+            0.125,
+            0.25,
+            0.53125
+        ],
+        "group_weights": [
+            0.2,
+            0.2,
+            0.2,
+            0.2,
+            0.2
+        ],
+        "group_sizes": [
+            2048,
+            4096,
+            8192,
+            16384,
+            34816
+        ],
+        "k": 130,
+        "device": "cuda:0",
+        "layer": 12,
+        "lm_name": "google/gemma-2-2b",
+        "wandb_name": "MatryoshkaBatchTopKTrainer-google/gemma-2-2b-resid_post_layer_12_trainer_7",
+        "submodule_name": "resid_post_layer_12"
+    },
+    "buffer": {
+        "d_submodule": 2304,
+        "io": "out",
+        "n_ctxs": 244,
+        "ctx_len": 1024,
+        "refresh_batch_size": 48,
+        "out_batch_size": 2048,
+        "device": "cuda:0"
+    }
+}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_7/eval_results.json

@@ -0,0 +1 @@
+{"l2_loss": 48.86486486486486, "l1_loss": 767.9639639639639, "l0": 130.0410169996657, "frac_variance_explained": 0.8726421734234234, "cossim": 0.94140625, "l2_ratio": 0.953125, "relative_reconstruction_bias": 1.0092905405405406, "loss_original": 2.319045608108108, "loss_reconstructed": 2.3657798423423424, "loss_zero": 12.4375, "frac_recovered": 0.9946157094594594, "frac_alive": 0.9301910400390625, "hyperparameters": {"n_inputs": 1000, "context_length": 1024}}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_8/config.json

@@ -0,0 +1,53 @@
+{
+    "trainer": {
+        "trainer_class": "MatryoshkaBatchTopKTrainer",
+        "dict_class": "MatryoshkaBatchTopKSAE",
+        "lr": 0.0002,
+        "steps": 244140,
+        "auxk_alpha": 0.03125,
+        "warmup_steps": 1000,
+        "decay_start": 195312,
+        "threshold_beta": 0.999,
+        "threshold_start_step": 1000,
+        "top_k_aux": 1152,
+        "seed": 721077,
+        "activation_dim": 2304,
+        "dict_size": 65536,
+        "group_fractions": [
+            0.03125,
+            0.0625,
+            0.125,
+            0.25,
+            0.53125
+        ],
+        "group_weights": [
+            0.2,
+            0.2,
+            0.2,
+            0.2,
+            0.2
+        ],
+        "group_sizes": [
+            2048,
+            4096,
+            8192,
+            16384,
+            34816
+        ],
+        "k": 25,
+        "device": "cuda:0",
+        "layer": 12,
+        "lm_name": "google/gemma-2-2b",
+        "wandb_name": "MatryoshkaBatchTopKTrainer-google/gemma-2-2b-resid_post_layer_12_trainer_8",
+        "submodule_name": "resid_post_layer_12"
+    },
+    "buffer": {
+        "d_submodule": 2304,
+        "io": "out",
+        "n_ctxs": 244,
+        "ctx_len": 1024,
+        "refresh_batch_size": 32,
+        "out_batch_size": 2048,
+        "device": "cuda:0"
+    }
+}

home/dontsov/interp/dictionary_learning/run_matr/google_gemma-2-2b_matryoshka_batch_top_k/resid_post_layer_12/trainer_9/config.json

@@ -0,0 +1,53 @@
+{
+    "trainer": {
+        "trainer_class": "MatryoshkaBatchTopKTrainer",
+        "dict_class": "MatryoshkaBatchTopKSAE",
+        "lr": 0.0002,
+        "steps": 244140,
+        "auxk_alpha": 0.03125,
+        "warmup_steps": 1000,
+        "decay_start": 195312,
+        "threshold_beta": 0.999,
+        "threshold_start_step": 1000,
+        "top_k_aux": 1152,
+        "seed": 721077,
+        "activation_dim": 2304,
+        "dict_size": 65536,
+        "group_fractions": [
+            0.03125,
+            0.0625,
+            0.125,
+            0.25,
+            0.53125
+        ],
+        "group_weights": [
+            0.2,
+            0.2,
+            0.2,
+            0.2,
+            0.2
+        ],
+        "group_sizes": [
+            2048,
+            4096,
+            8192,
+            16384,
+            34816
+        ],
+        "k": 40,
+        "device": "cuda:0",
+        "layer": 12,
+        "lm_name": "google/gemma-2-2b",
+        "wandb_name": "MatryoshkaBatchTopKTrainer-google/gemma-2-2b-resid_post_layer_12_trainer_9",
+        "submodule_name": "resid_post_layer_12"
+    },
+    "buffer": {
+        "d_submodule": 2304,
+        "io": "out",
+        "n_ctxs": 244,
+        "ctx_len": 1024,
+        "refresh_batch_size": 32,
+        "out_batch_size": 2048,
+        "device": "cuda:0"
+    }
+}