Upload CrossCoder config for layer blocks.14.hook_resid_pre (step N/A)
Browse files
blocks.14.hook_resid_pre/cfg.json
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"seed": 49,
|
| 3 |
+
"device": "cuda:0",
|
| 4 |
+
"wandb_project": "gemma-crosscoder",
|
| 5 |
+
"wandb_entity": "kasmik-backups-kathmandu-university",
|
| 6 |
+
"wandb_run_name": null,
|
| 7 |
+
"checkpoint_path": "./checkpoints/",
|
| 8 |
+
"data_cache_dir": "./cache/",
|
| 9 |
+
"data_dir": "./data/",
|
| 10 |
+
"model_name_A": "google/gemma-2-2b",
|
| 11 |
+
"model_name_B": "google/gemma-2-2b-it",
|
| 12 |
+
"hook_point": "blocks.14.hook_resid_pre",
|
| 13 |
+
"dict_size": 16384,
|
| 14 |
+
"enc_dtype": "fp32",
|
| 15 |
+
"dec_init_norm": 0.08,
|
| 16 |
+
"lr": 5e-05,
|
| 17 |
+
"beta1": 0.9,
|
| 18 |
+
"beta2": 0.999,
|
| 19 |
+
"l1_coeff": 2.0,
|
| 20 |
+
"lr_warmup_steps": 0,
|
| 21 |
+
"lr_decay_start_fraction": 0.8,
|
| 22 |
+
"l1_warmup_fraction": 0.05,
|
| 23 |
+
"num_tokens": 10000000,
|
| 24 |
+
"batch_size": 4096,
|
| 25 |
+
"seq_len": 1024,
|
| 26 |
+
"model_batch_size": 4,
|
| 27 |
+
"buffer_mult": 128,
|
| 28 |
+
"norm_batch_size": 16,
|
| 29 |
+
"log_every": 100,
|
| 30 |
+
"save_every": 20000,
|
| 31 |
+
"d_in": 2304
|
| 32 |
+
}
|