Upload Llama-3.2-1B-L8-k100-lr1e-04-local-shuffling-Crosscoder dictionary model
Browse files- trainer_config.json +35 -0
trainer_config.json
ADDED
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"trainer": {
|
3 |
+
"trainer_class": "BatchTopKCrossCoderTrainer",
|
4 |
+
"dict_class": "BatchTopKCrossCoder",
|
5 |
+
"lr": 0.0001,
|
6 |
+
"steps": 97656,
|
7 |
+
"auxk_alpha": 0.03125,
|
8 |
+
"warmup_steps": 1000,
|
9 |
+
"decay_start": null,
|
10 |
+
"threshold_beta": 0.999,
|
11 |
+
"threshold_start_step": 1000,
|
12 |
+
"top_k_aux": 1024,
|
13 |
+
"seed": null,
|
14 |
+
"activation_dim": 2048,
|
15 |
+
"dict_size": 65536,
|
16 |
+
"k": 100,
|
17 |
+
"code_normalization": "CROSSCODER",
|
18 |
+
"code_normalization_alpha_sae": 1.0,
|
19 |
+
"code_normalization_alpha_cc": 0.1,
|
20 |
+
"device": "cuda",
|
21 |
+
"layer": 8,
|
22 |
+
"lm_name": "meta-llama/Llama-3.2-1B-Instruct-meta-llama/Llama-3.2-1B",
|
23 |
+
"wandb_name": "Llama-3.2-1B-L8-k100-lr1e-04-local-shuffling-Crosscoder",
|
24 |
+
"submodule_name": null,
|
25 |
+
"dict_class_kwargs": {
|
26 |
+
"same_init_for_all_layers": "True",
|
27 |
+
"norm_init_scale": "1.0",
|
28 |
+
"init_with_transpose": "True",
|
29 |
+
"encoder_layers": "None",
|
30 |
+
"code_normalization": "crosscoder",
|
31 |
+
"code_normalization_alpha_sae": "1.0",
|
32 |
+
"code_normalization_alpha_cc": "0.1"
|
33 |
+
}
|
34 |
+
}
|
35 |
+
}
|