Upload Meta-Llama-3.1-8B-L16-k200-lr1e-04-local-shuffling-Crosscoder-ni0.3-ka1k5k dictionary model
Browse files- trainer_config.json +38 -0
trainer_config.json
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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": 2048,
|
| 13 |
+
"seed": null,
|
| 14 |
+
"activation_dim": 4096,
|
| 15 |
+
"dict_size": 131072,
|
| 16 |
+
"k": 1000,
|
| 17 |
+
"k_target": 200,
|
| 18 |
+
"k_initial": 1000,
|
| 19 |
+
"k_annealing_steps": 5000,
|
| 20 |
+
"code_normalization": "CROSSCODER",
|
| 21 |
+
"code_normalization_alpha_sae": 1.0,
|
| 22 |
+
"code_normalization_alpha_cc": 0.1,
|
| 23 |
+
"device": "cuda",
|
| 24 |
+
"layer": 16,
|
| 25 |
+
"lm_name": "meta-llama/Meta-Llama-3.1-8B-Instruct-meta-llama/Meta-Llama-3.1-8B",
|
| 26 |
+
"wandb_name": "Meta-Llama-3.1-8B-L16-k200-lr1e-04-local-shuffling-Crosscoder-ni0.3-ka1k5k",
|
| 27 |
+
"submodule_name": null,
|
| 28 |
+
"dict_class_kwargs": {
|
| 29 |
+
"same_init_for_all_layers": "True",
|
| 30 |
+
"norm_init_scale": "0.3",
|
| 31 |
+
"init_with_transpose": "True",
|
| 32 |
+
"encoder_layers": "None",
|
| 33 |
+
"code_normalization": "crosscoder",
|
| 34 |
+
"code_normalization_alpha_sae": "1.0",
|
| 35 |
+
"code_normalization_alpha_cc": "0.1"
|
| 36 |
+
}
|
| 37 |
+
}
|
| 38 |
+
}
|