Upload folder using huggingface_hub

gemma_scope/core_1M/gemma-scope-2b-pt-res_layer_12_width_1m_average_l0_19_eval_results.json

@@ -0,0 +1,87 @@
+{
+  "eval_type_id": "core",
+  "eval_config": {
+    "model_name": "gemma-2-2b",
+    "llm_dtype": "bfloat16",
+    "batch_size_prompts": 2,
+    "n_eval_reconstruction_batches": 1600,
+    "n_eval_sparsity_variance_batches": 16000,
+    "dataset": "Skylion007/openwebtext",
+    "context_size": 128,
+    "compute_kl": true,
+    "compute_ce_loss": true,
+    "compute_l2_norms": true,
+    "compute_sparsity_metrics": true,
+    "compute_variance_metrics": true,
+    "compute_featurewise_density_statistics": false,
+    "compute_featurewise_weight_based_metrics": false,
+    "exclude_special_tokens_from_reconstruction": true,
+    "verbose": false
+  },
+  "eval_id": "a4041870-141a-4321-9012-28d11da1b3cc",
+  "datetime_epoch_millis": 1733113515858,
+  "eval_result_metrics": {
+    "model_behavior_preservation": {
+      "kl_div_score": 0.983792701863354,
+      "kl_div_with_ablation": 10.0625,
+      "kl_div_with_sae": 0.1630859375
+    },
+    "model_performance_preservation": {
+      "ce_loss_score": 0.9835526315789473,
+      "ce_loss_with_ablation": 12.4375,
+      "ce_loss_with_sae": 3.09375,
+      "ce_loss_without_sae": 2.9375
+    },
+    "reconstruction_quality": {
+      "explained_variance": 0.70703125,
+      "mse": 13.75,
+      "cossim": 0.9140625
+    },
+    "shrinkage": {
+      "l2_norm_in": 149.0,
+      "l2_norm_out": 137.0,
+      "l2_ratio": 0.9140625,
+      "relative_reconstruction_bias": 1.0
+    },
+    "sparsity": {
+      "l0": 18.4305419921875,
+      "l1": 392.0
+    },
+    "token_stats": {
+      "total_tokens_eval_reconstruction": 409600,
+      "total_tokens_eval_sparsity_variance": 4096000
+    }
+  },
+  "eval_result_details": [],
+  "sae_bench_commit_hash": "67c9b0381e8bbf0ef5a845dee64aa2b955a47213",
+  "sae_lens_id": "layer_12/width_1m/average_l0_19",
+  "sae_lens_release_id": "gemma-scope-2b-pt-res",
+  "sae_lens_version": "4.4.5",
+  "sae_cfg_dict": {
+    "architecture": "jumprelu",
+    "d_in": 2304,
+    "d_sae": 1048576,
+    "activation_fn_str": "relu",
+    "apply_b_dec_to_input": false,
+    "finetuning_scaling_factor": false,
+    "context_size": 1024,
+    "model_name": "gemma-2-2b",
+    "hook_name": "blocks.12.hook_resid_post",
+    "hook_layer": 12,
+    "hook_head_index": null,
+    "prepend_bos": true,
+    "dataset_path": "monology/pile-uncopyrighted",
+    "dataset_trust_remote_code": true,
+    "normalize_activations": null,
+    "dtype": "torch.bfloat16",
+    "device": "cuda",
+    "sae_lens_training_version": null,
+    "activation_fn_kwargs": {},
+    "neuronpedia_id": "gemma-2-2b/12-gemmascope-res-1m__l0-19",
+    "model_from_pretrained_kwargs": {},
+    "seqpos_slice": [
+      null
+    ]
+  },
+  "eval_result_unstructured": {}
+}

gemma_scope/core_1M/gemma-scope-2b-pt-res_layer_12_width_1m_average_l0_207_eval_results.json

@@ -0,0 +1,87 @@
+{
+  "eval_type_id": "core",
+  "eval_config": {
+    "model_name": "gemma-2-2b",
+    "llm_dtype": "bfloat16",
+    "batch_size_prompts": 2,
+    "n_eval_reconstruction_batches": 1600,
+    "n_eval_sparsity_variance_batches": 16000,
+    "dataset": "Skylion007/openwebtext",
+    "context_size": 128,
+    "compute_kl": true,
+    "compute_ce_loss": true,
+    "compute_l2_norms": true,
+    "compute_sparsity_metrics": true,
+    "compute_variance_metrics": true,
+    "compute_featurewise_density_statistics": false,
+    "compute_featurewise_weight_based_metrics": false,
+    "exclude_special_tokens_from_reconstruction": true,
+    "verbose": false
+  },
+  "eval_id": "3e5b376c-bd18-4ff7-82ab-dd0e5d4ffec7",
+  "datetime_epoch_millis": 1733115669696,
+  "eval_result_metrics": {
+    "model_behavior_preservation": {
+      "kl_div_score": 0.9970035908385093,
+      "kl_div_with_ablation": 10.0625,
+      "kl_div_with_sae": 0.0301513671875
+    },
+    "model_performance_preservation": {
+      "ce_loss_score": 0.9983552631578947,
+      "ce_loss_with_ablation": 12.4375,
+      "ce_loss_with_sae": 2.953125,
+      "ce_loss_without_sae": 2.9375
+    },
+    "reconstruction_quality": {
+      "explained_variance": 0.875,
+      "mse": 5.875,
+      "cossim": 0.96484375
+    },
+    "shrinkage": {
+      "l2_norm_in": 149.0,
+      "l2_norm_out": 144.0,
+      "l2_ratio": 0.9609375,
+      "relative_reconstruction_bias": 1.0
+    },
+    "sparsity": {
+      "l0": 210.41099548339844,
+      "l1": 756.0
+    },
+    "token_stats": {
+      "total_tokens_eval_reconstruction": 409600,
+      "total_tokens_eval_sparsity_variance": 4096000
+    }
+  },
+  "eval_result_details": [],
+  "sae_bench_commit_hash": "67c9b0381e8bbf0ef5a845dee64aa2b955a47213",
+  "sae_lens_id": "layer_12/width_1m/average_l0_207",
+  "sae_lens_release_id": "gemma-scope-2b-pt-res",
+  "sae_lens_version": "4.4.5",
+  "sae_cfg_dict": {
+    "architecture": "jumprelu",
+    "d_in": 2304,
+    "d_sae": 1048576,
+    "activation_fn_str": "relu",
+    "apply_b_dec_to_input": false,
+    "finetuning_scaling_factor": false,
+    "context_size": 1024,
+    "model_name": "gemma-2-2b",
+    "hook_name": "blocks.12.hook_resid_post",
+    "hook_layer": 12,
+    "hook_head_index": null,
+    "prepend_bos": true,
+    "dataset_path": "monology/pile-uncopyrighted",
+    "dataset_trust_remote_code": true,
+    "normalize_activations": null,
+    "dtype": "torch.bfloat16",
+    "device": "cuda:0",
+    "sae_lens_training_version": null,
+    "activation_fn_kwargs": {},
+    "neuronpedia_id": "gemma-2-2b/12-gemmascope-res-1m__l0-207",
+    "model_from_pretrained_kwargs": {},
+    "seqpos_slice": [
+      null
+    ]
+  },
+  "eval_result_unstructured": {}
+}

gemma_scope/core_1M/gemma-scope-2b-pt-res_layer_5_width_1m_average_l0_9_eval_results.json

@@ -0,0 +1,87 @@
+{
+  "eval_type_id": "core",
+  "eval_config": {
+    "model_name": "gemma-2-2b",
+    "llm_dtype": "bfloat16",
+    "batch_size_prompts": 2,
+    "n_eval_reconstruction_batches": 1600,
+    "n_eval_sparsity_variance_batches": 16000,
+    "dataset": "Skylion007/openwebtext",
+    "context_size": 128,
+    "compute_kl": true,
+    "compute_ce_loss": true,
+    "compute_l2_norms": true,
+    "compute_sparsity_metrics": true,
+    "compute_variance_metrics": true,
+    "compute_featurewise_density_statistics": false,
+    "compute_featurewise_weight_based_metrics": false,
+    "exclude_special_tokens_from_reconstruction": true,
+    "verbose": false
+  },
+  "eval_id": "3189e07f-932c-4881-86e7-ed01fa6ceb03",
+  "datetime_epoch_millis": 1733111254164,
+  "eval_result_metrics": {
+    "model_behavior_preservation": {
+      "kl_div_score": 0.9933035714285714,
+      "kl_div_with_ablation": 10.0625,
+      "kl_div_with_sae": 0.0673828125
+    },
+    "model_performance_preservation": {
+      "ce_loss_score": 0.9950657894736842,
+      "ce_loss_with_ablation": 12.4375,
+      "ce_loss_with_sae": 2.984375,
+      "ce_loss_without_sae": 2.9375
+    },
+    "reconstruction_quality": {
+      "explained_variance": 0.83203125,
+      "mse": 3.546875,
+      "cossim": 0.9453125
+    },
+    "shrinkage": {
+      "l2_norm_in": 90.0,
+      "l2_norm_out": 84.0,
+      "l2_ratio": 0.9375,
+      "relative_reconstruction_bias": 0.99609375
+    },
+    "sparsity": {
+      "l0": 9.345510482788086,
+      "l1": 140.0
+    },
+    "token_stats": {
+      "total_tokens_eval_reconstruction": 409600,
+      "total_tokens_eval_sparsity_variance": 4096000
+    }
+  },
+  "eval_result_details": [],
+  "sae_bench_commit_hash": "67c9b0381e8bbf0ef5a845dee64aa2b955a47213",
+  "sae_lens_id": "layer_5/width_1m/average_l0_9",
+  "sae_lens_release_id": "gemma-scope-2b-pt-res",
+  "sae_lens_version": "4.4.5",
+  "sae_cfg_dict": {
+    "architecture": "jumprelu",
+    "d_in": 2304,
+    "d_sae": 1048576,
+    "activation_fn_str": "relu",
+    "apply_b_dec_to_input": false,
+    "finetuning_scaling_factor": false,
+    "context_size": 1024,
+    "model_name": "gemma-2-2b",
+    "hook_name": "blocks.5.hook_resid_post",
+    "hook_layer": 5,
+    "hook_head_index": null,
+    "prepend_bos": true,
+    "dataset_path": "monology/pile-uncopyrighted",
+    "dataset_trust_remote_code": true,
+    "normalize_activations": null,
+    "dtype": "torch.bfloat16",
+    "device": "cuda",
+    "sae_lens_training_version": null,
+    "activation_fn_kwargs": {},
+    "neuronpedia_id": "gemma-2-2b/5-gemmascope-res-1m__l0-9",
+    "model_from_pretrained_kwargs": {},
+    "seqpos_slice": [
+      null
+    ]
+  },
+  "eval_result_unstructured": {}
+}