alexquach commited on
Commit
4098f0c
·
verified ·
1 Parent(s): 89c0af6

Upload folder using huggingface_hub

Browse files
._fast_run_4096w_40k_256bz_s1_l40s_google_gemma-2-2b_nested_dropout/resid_post_layer_12/trainer_0/ae.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a2d2c2a1c511333475b97e09dbecbf316706eae3140bac9e83cbf5f383daa300
3
+ size 75525078
._fast_run_4096w_40k_256bz_s1_l40s_google_gemma-2-2b_nested_dropout/resid_post_layer_12/trainer_0/config.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "trainer": {
3
+ "trainer_class": "NestedDropoutBatchTopKTrainer",
4
+ "dict_class": "NestedDropoutBatchTopKSAE",
5
+ "lr": 0.0003,
6
+ "steps": 195312,
7
+ "auxk_alpha": 0.03125,
8
+ "warmup_steps": 1000,
9
+ "decay_start": 156249,
10
+ "threshold_beta": 0.999,
11
+ "threshold_start_step": 1000,
12
+ "top_k_aux": 1152,
13
+ "seed": 1,
14
+ "activation_dim": 2304,
15
+ "dict_size": 4096,
16
+ "p_L_over_p_1": 0.001,
17
+ "k": 40,
18
+ "device": "cuda",
19
+ "layer": 12,
20
+ "lm_name": "google/gemma-2-2b",
21
+ "wandb_name": "ND-4096-40-1_trainer_0",
22
+ "submodule_name": "resid_post_layer_12"
23
+ },
24
+ "buffer": {
25
+ "d_submodule": 2304,
26
+ "io": "out",
27
+ "n_ctxs": 264,
28
+ "ctx_len": 1024,
29
+ "refresh_batch_size": 4,
30
+ "out_batch_size": 256,
31
+ "device": "cuda:0"
32
+ }
33
+ }
._fast_run_4096w_40k_256bz_s1_l40s_google_gemma-2-2b_nested_dropout/resid_post_layer_12/trainer_0/eval_results.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"l2_loss": 79.93, "l1_loss": 514.86, "l0": 39.8348095703125, "frac_variance_explained": 0.68361328125, "cossim": 0.82630859375, "l2_ratio": 0.8315625, "relative_reconstruction_bias": 1.00609375, "loss_original": 2.152509765625, "loss_reconstructed": 3.0796875, "loss_zero": 12.4375, "frac_recovered": 0.909609375, "frac_alive": 0.996826171875, "hyperparameters": {"n_inputs": 200, "context_length": 1024}}