Upload GPT2_L1_lr_0.0001_transformer.h.1.attn_config.json with huggingface_hub
Browse files
GPT2_L1_lr_0.0001_transformer.h.1.attn_config.json
ADDED
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"seed": 42,
|
3 |
+
"model_name_or_path": "gpt2",
|
4 |
+
"revision": "",
|
5 |
+
"hook_point": "transformer.h.1.attn",
|
6 |
+
"dataset_name_or_path": "jbrinkma/pile-100k",
|
7 |
+
"activation_size": -1,
|
8 |
+
"add_bos_token": false,
|
9 |
+
"evaluation_batches": 10,
|
10 |
+
"expansion_factor": 4,
|
11 |
+
"b_dec_init_method": "",
|
12 |
+
"use_pre_encoder_bias": true,
|
13 |
+
"tied": false,
|
14 |
+
"n_steps": -1,
|
15 |
+
"device": "cuda",
|
16 |
+
"batch_size": 32,
|
17 |
+
"ctx_length": 256,
|
18 |
+
"lr": 0.0001,
|
19 |
+
"min_lr": 0.0,
|
20 |
+
"lr_warmup_steps": 5000,
|
21 |
+
"sparsity_coefficient": 0.003,
|
22 |
+
"evaluation_interval": 200,
|
23 |
+
"beta1": 0.9,
|
24 |
+
"beta2": 0.999,
|
25 |
+
"l1_sqrt": false,
|
26 |
+
"cos_sim_reg": false,
|
27 |
+
"cos_sim_alpha": 0.0,
|
28 |
+
"decoder_normalization": true,
|
29 |
+
"decoder_norm_smaller_than_one": false,
|
30 |
+
"l1_with_norm": false,
|
31 |
+
"sqrt_mse": false,
|
32 |
+
"dynamic_weighting": false,
|
33 |
+
"l1_warmup_steps": 2000,
|
34 |
+
"target_l0": 40,
|
35 |
+
"n_tokens_in_feature_cache": 500000.0,
|
36 |
+
"use_ghost_grads": false,
|
37 |
+
"use_neuron_resampling": false,
|
38 |
+
"resampling_steps": -1,
|
39 |
+
"output_dir": "outputs",
|
40 |
+
"cache_dir": "cache",
|
41 |
+
"checkpoint_interval": 200,
|
42 |
+
"use_wandb": true,
|
43 |
+
"wandb_entity": "jannikbrinkmann",
|
44 |
+
"wandb_project": "best-sae",
|
45 |
+
"wandb_name": "GPT2_L1_lr_0.0001_20240323115650955927",
|
46 |
+
"wandb_group": "L0_L1_Sweeps"
|
47 |
+
}
|