Upload SAE blocks.1.hook_resid_post/l0_12/ortho_10k/cutoff_0.02/threshold_0.5

blocks.1.hook_resid_post/l0_12/ortho_10k/cutoff_0.02/threshold_0.5/cfg.json

@@ -0,0 +1 @@
+{"model_name": "gemma-2-2b", "model_class_name": "HookedTransformer", "hook_name": "blocks.1.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 1, "hook_head_index": null, "dataset_path": "HuggingFaceFW/fineweb", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 2304, "d_sae": 18432, "b_dec_init_method": "geometric_median", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 48, "training_tokens": 1000000000, "finetuning_tokens": 0, "store_batch_size_prompts": 6, "train_batch_size_tokens": 4096, "normalize_activations": "none", "seqpos_slice": [null], "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "jumprelu_init_threshold": 0.001, "jumprelu_bandwidth": 0.001, "autocast": true, "autocast_lm": true, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 0.5, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 15000, "lr": 2e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 2.0000000000000003e-06, "lr_decay_steps": 0, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-08, "n_eval_batches": 10, "eval_batch_size_prompts": 1, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "gemma-2-2b-layer-1", "wandb_id": null, "run_name": "gemma-2-2b-tied-act-ortho-layer-1-1000000000-oc-10000.0-co-0.02-l1-0.5-th-0.5-wu-10000-w-50-2024-12-07T03:55:48", "wandb_entity": "anti-absorb", "wandb_log_frequency": 100, "eval_every_n_wandb_logs": 50, "resume": false, "n_checkpoints": 1, "checkpoint_path": "/home/ucabdj4/Scratch/sae-gemma-2-2b-tied-act-ortho-tl-0.5-l1-0.5-oc-1e4-co-0.02-ri-30k-mr-4-rt-0.3-wu-10k-w-50-checkpoints/kbrdnv3q", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {"center_writing_weights": false}, "sae_lens_version": "4.4.5", "sae_lens_training_version": "4.4.5", "ortho_coefficient": 10000.0, "ortho_cutoff": 0.02, "max_magnitudes_window_size_batches": 50, "ortho_loss_warm_up_steps": 10000, "add_ortho_loss_after_n_steps": 50000, "ortho_reinit_interval_steps": 30000, "max_ortho_reinits": 4, "ortho_reinit_cos_threshold": 0.3, "remove_l1_after_warmup": true, "threshold_level": 0.5, "tokens_per_buffer": 201326592, "activation_fn_str": "relu", "finetuning_scaling_factor": false, "neuronpedia_id": null}

blocks.1.hook_resid_post/l0_12/ortho_10k/cutoff_0.02/threshold_0.5/sae_weights.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7325723a650700f6328574cd247ec65fe84bcf3e6302d28b6d3f353a7c686498
+size 339895696

blocks.1.hook_resid_post/l0_12/ortho_10k/cutoff_0.02/threshold_0.5/sparsity.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55de3bb2f4589540bf294b8b5cc776218d4c60d3ebd1e74a59d19c0ff5e2b23a
+size 73808