File size: 2,263 Bytes
8c1164c |
1 |
{"model_name": "gpt2", "model_class_name": "HookedTransformer", "hook_point": "blocks.10.hook_attn_out", "hook_point_eval": "blocks.{layer}.attn.pattern", "hook_point_layer": 10, "hook_point_head_index": null, "dataset_path": "apollo-research/Skylion007-openwebtext-tokenizer-gpt2", "streaming": true, "is_dataset_tokenized": true, "context_size": 128, "use_cached_activations": false, "cached_activations_path": null, "d_in": 768, "d_sae": 49152, "b_dec_init_method": "zeros", "expansion_factor": 64, "activation_fn": "relu", "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 128, "training_tokens": 61440000, "finetuning_tokens": 0, "store_batch_size_prompts": 32, "train_batch_size_tokens": 4096, "normalize_activations": false, "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "torch.float32", "prepend_bos": true, "autocast": true, "autocast_lm": false, "compile_llm": true, "llm_compilation_mode": null, "compile_sae": true, "sae_compilation_mode": null, "adam_beta1": 0.9, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1.0, "scale_sparsity_penalty_by_decoder_norm": true, "l1_warm_up_steps": 1500, "lr": 5e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 5e-06, "lr_decay_steps": 0, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 1000, "dead_feature_window": 1000, "dead_feature_threshold": 0.0001, "n_eval_batches": 10, "eval_batch_size_prompts": 2, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "transfer-learning", "wandb_id": null, "run_name": "49152-L1-1-LR-5e-05-Tokens-6.144e+07", "wandb_entity": null, "wandb_log_frequency": 50, "eval_every_n_wandb_logs": 10, "resume": false, "n_checkpoints": 0, "checkpoint_path": "sae-transfer-learning/first-attn-transfer-gpt2/e37amtur", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "2.1.3", "sae_lens_training_version": "2.1.3", "tokens_per_buffer": 67108864} |