Hzfinfdu commited on
Commit
afeb0ec
1 Parent(s): fe37936

Rename folders

Browse files
Llama3_1-8B-Base-L20M-8x/hyperparams.json ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "device": "cuda:0",
3
+ "seed": 42,
4
+ "dtype": "torch.bfloat16",
5
+ "hook_point_in": "blocks.20.hook_mlp_out",
6
+ "hook_point_out": "blocks.20.hook_mlp_out",
7
+ "use_decoder_bias": true,
8
+ "apply_decoder_bias_to_pre_encoder": false,
9
+ "expansion_factor": 8,
10
+ "d_model": 4096,
11
+ "d_sae": 32768,
12
+ "bias_init_method": "all_zero",
13
+ "act_fn": "jumprelu",
14
+ "jump_relu_threshold": 0.2109375,
15
+ "norm_activation": "dataset-wise",
16
+ "dataset_average_activation_norm": {
17
+ "in": 6.3125,
18
+ "out": 6.3125
19
+ },
20
+ "decoder_exactly_fixed_norm": false,
21
+ "sparsity_include_decoder_norm": true,
22
+ "use_glu_encoder": false,
23
+ "init_decoder_norm": 0.5,
24
+ "init_encoder_norm": null,
25
+ "init_encoder_with_decoder_transpose": true,
26
+ "lp": 1,
27
+ "l1_coefficient": 8e-05,
28
+ "l1_coefficient_warmup_steps": 39062,
29
+ "top_k": 50,
30
+ "k_warmup_steps": 39062,
31
+ "use_batch_norm_mse": true,
32
+ "use_ghost_grads": false,
33
+ "tp_size": 1,
34
+ "ddp_size": 1
35
+ }