File size: 1,245 Bytes
b2797fb
 
 
 
 
 
 
 
 
 
144fd82
 
b2797fb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
144fd82
 
b2797fb
144fd82
 
 
 
b2797fb
 
144fd82
 
 
 
b2797fb
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
{
  "architecture": "cross_model_crosscoder",
  "variant": "BatchTopK with JumpReLU inference threshold",
  "reference": "Lindsey 2024 + Anthropic Jan-2025 update + Minder NeurIPS 2025",
  "base_model": "google/gemma-2-2b",
  "chat_model": "google/gemma-2-2b-it",
  "layer": 13,
  "d_model": 2304,
  "expansion": 32,
  "k_batchtopk": 100,
  "k_warmup_init": 1000,
  "k_warmup_steps": 5000,
  "dec_init_norm": 1.0,
  "token_budget": 100000000,
  "seq_len": 512,
  "fwd_batch": 4,
  "cc_batch": 4096,
  "lr": 0.0001,
  "lambda_l1": 0.041,
  "warmup_steps": 1000,
  "grad_clip": 1.0,
  "lr_decay_frac": 0.2,
  "checkpoint_every_tokens": 5000000,
  "data_web_frac": 0.5,
  "data_chat_frac": 0.5,
  "hf_user": "caiovicentino1",
  "hf_repo_name": "gemma2-2b-crosscoder-model-diff-papergrade",
  "n_features": 73728,
  "hf_repo": "caiovicentino1/gemma2-2b-crosscoder-model-diff-papergrade",
  "norm_scale_A": 0.2677689692703698,
  "norm_scale_B": 0.23634909321590494,
  "val_metrics": {
    "ve_A": 0.8773096382617951,
    "ve_B": 0.8665769904851913,
    "L0": 100.53505859375,
    "dead_frac": 0.4289415180683136
  },
  "taxonomy_counts": {
    "shared": 39711,
    "dead": 31625,
    "unclassified": 2385,
    "base_only": 4,
    "chat_only": 3
  }
}