Initial commit
Browse files
ancient-feather-33/hyperparameters.yaml
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
adam_beta1: 0.9
|
| 2 |
+
adam_beta2: 0.999
|
| 3 |
+
b_dec_init_method: zeros
|
| 4 |
+
cached_activations_path: null
|
| 5 |
+
checkpoint_path: ./outputs/checkpoints
|
| 6 |
+
clip_grad_norm: true
|
| 7 |
+
context_size: 256
|
| 8 |
+
custom_loss: null
|
| 9 |
+
d_in: 2560
|
| 10 |
+
d_out: null
|
| 11 |
+
dataset: monology/pile-uncopyrighted
|
| 12 |
+
dense_loss_coefficient: 0
|
| 13 |
+
device: cuda
|
| 14 |
+
different_output: false
|
| 15 |
+
dtype: float32
|
| 16 |
+
epsilon_l0_approx: 0.5
|
| 17 |
+
eval_frequency: 500
|
| 18 |
+
expansion_factor: 8
|
| 19 |
+
feature_reinit_scale: 0.2
|
| 20 |
+
feature_resampling_method: null
|
| 21 |
+
fine_tune_dataset: false
|
| 22 |
+
flatten_activations_over_layer: false
|
| 23 |
+
flatten_activations_over_layer_output: false
|
| 24 |
+
from_pretrained_path: null
|
| 25 |
+
hook_point: blocks.5.hook_resid_pre
|
| 26 |
+
hook_point_head_index: null
|
| 27 |
+
hook_point_head_index_output: null
|
| 28 |
+
hook_point_layer: 5
|
| 29 |
+
hook_point_layer_output: null
|
| 30 |
+
hook_point_output: null
|
| 31 |
+
initial_decoder_norm: 0.1
|
| 32 |
+
initialise_encoder_to_decoder_transpose: false
|
| 33 |
+
is_dataset_tokenized: false
|
| 34 |
+
l0_coefficient: 0
|
| 35 |
+
l0_warmup: false
|
| 36 |
+
l0_warmup_steps: 1000
|
| 37 |
+
l1_coefficient: 5
|
| 38 |
+
l1_warmup: true
|
| 39 |
+
l1_warmup_steps: 10000
|
| 40 |
+
log_to_wandb: true
|
| 41 |
+
loop_dataset: false
|
| 42 |
+
lr: 0.0001
|
| 43 |
+
lr_scheduler_name: constant_with_warmup
|
| 44 |
+
lr_warm_up_steps: 1000
|
| 45 |
+
max_resample_step: 100000
|
| 46 |
+
max_sparsity_target: 1
|
| 47 |
+
min_sparsity_for_resample: 0
|
| 48 |
+
min_sparsity_target: 0
|
| 49 |
+
model_name: EleutherAI/pythia-2.8b-deduped
|
| 50 |
+
mse_loss_coefficient: 1
|
| 51 |
+
mse_loss_type: standard
|
| 52 |
+
multiple_runs: false
|
| 53 |
+
n_batches_in_store_buffer: 128
|
| 54 |
+
n_checkpoints: 80
|
| 55 |
+
n_running_sparsity: 100
|
| 56 |
+
normalise_initial_decoder_weights: false
|
| 57 |
+
normalise_w_dec: false
|
| 58 |
+
resample_batches: 128
|
| 59 |
+
resample_frequency: 25000
|
| 60 |
+
scale_input_norm: false
|
| 61 |
+
seed: 42
|
| 62 |
+
sparse_loss_coefficient: 0
|
| 63 |
+
sparsity_log_frequency: 5000
|
| 64 |
+
store_batch_size: 8
|
| 65 |
+
subtract_b_dec_from_inputs: false
|
| 66 |
+
total_training_steps: 24414
|
| 67 |
+
train_batch_size: 4096
|
| 68 |
+
use_cached_activations: false
|
| 69 |
+
use_gated_sparse_autoencoder: false
|
| 70 |
+
wandb_log_frequency: 10
|
| 71 |
+
wandb_project: test_pythia-mlp
|
| 72 |
+
weight_l1_by_decoder_norms: true
|
ancient-feather-33/sparse_autoencoder/pythia-2.8b-deduped_blocks.0.hook_mlp_out_s5120_100M_tokens_thepile_L0-23.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:45cb2f38f2f80dfc1d2655dd30cbe8a133cdc95296a657a086723c8eb19d8054
|
| 3 |
+
size 419527664
|