Upload folder using huggingface_hub
Browse files- config.json +39 -0
- direction.pt +3 -0
- direction_metadata.json +4 -0
- generate_directions/mean_diffs.pt +3 -0
- merged_model.safetensors +3 -0
- select_direction/ablation_scores.png +0 -0
- select_direction/actadd_scores.png +0 -0
- select_direction/direction_evaluations.json +1122 -0
- select_direction/direction_evaluations_filtered.json +128 -0
- select_direction/kl_div_scores.png +0 -0
config.json
ADDED
|
@@ -0,0 +1,39 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_name_or_path": "meta-llama/Meta-Llama-3.1-8B-Instruct",
|
| 3 |
+
"architectures": [
|
| 4 |
+
"LlamaForCausalLM"
|
| 5 |
+
],
|
| 6 |
+
"attention_bias": false,
|
| 7 |
+
"attention_dropout": 0.0,
|
| 8 |
+
"bos_token_id": 128000,
|
| 9 |
+
"eos_token_id": [
|
| 10 |
+
128001,
|
| 11 |
+
128008,
|
| 12 |
+
128009
|
| 13 |
+
],
|
| 14 |
+
"hidden_act": "silu",
|
| 15 |
+
"hidden_size": 4096,
|
| 16 |
+
"initializer_range": 0.02,
|
| 17 |
+
"intermediate_size": 14336,
|
| 18 |
+
"max_position_embeddings": 131072,
|
| 19 |
+
"mlp_bias": false,
|
| 20 |
+
"model_type": "llama",
|
| 21 |
+
"num_attention_heads": 32,
|
| 22 |
+
"num_hidden_layers": 32,
|
| 23 |
+
"num_key_value_heads": 8,
|
| 24 |
+
"pretraining_tp": 1,
|
| 25 |
+
"rms_norm_eps": 1e-05,
|
| 26 |
+
"rope_scaling": {
|
| 27 |
+
"factor": 8.0,
|
| 28 |
+
"high_freq_factor": 4.0,
|
| 29 |
+
"low_freq_factor": 1.0,
|
| 30 |
+
"original_max_position_embeddings": 8192,
|
| 31 |
+
"rope_type": "llama3"
|
| 32 |
+
},
|
| 33 |
+
"rope_theta": 500000.0,
|
| 34 |
+
"tie_word_embeddings": false,
|
| 35 |
+
"torch_dtype": "bfloat16",
|
| 36 |
+
"transformers_version": "4.43.3",
|
| 37 |
+
"use_cache": true,
|
| 38 |
+
"vocab_size": 128256
|
| 39 |
+
}
|
direction.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c03b0ade42731c2c5e87c145db111310c405effea28221b1be47716876ac864f
|
| 3 |
+
size 5244070
|
direction_metadata.json
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"pos": -2,
|
| 3 |
+
"layer": 12
|
| 4 |
+
}
|
generate_directions/mean_diffs.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5675bf91ace106199ca3c6e0376fe66e541beef6e3eead2d140c004abc734f5a
|
| 3 |
+
size 5244075
|
merged_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:40d95b3683b2e1ca91a9e854553003bd023cb774e0f37c1730e223c0c381c9a5
|
| 3 |
+
size 32121079000
|
select_direction/ablation_scores.png
ADDED
|
select_direction/actadd_scores.png
ADDED
|
select_direction/direction_evaluations.json
ADDED
|
@@ -0,0 +1,1122 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"position": -5,
|
| 4 |
+
"layer": 0,
|
| 5 |
+
"refusal_score": 7.450636863708496,
|
| 6 |
+
"steering_score": -11.830456733703613,
|
| 7 |
+
"kl_div_score": 0.0
|
| 8 |
+
},
|
| 9 |
+
{
|
| 10 |
+
"position": -5,
|
| 11 |
+
"layer": 1,
|
| 12 |
+
"refusal_score": 10.940382957458496,
|
| 13 |
+
"steering_score": -11.839740753173828,
|
| 14 |
+
"kl_div_score": 0.25063224011473184
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"position": -5,
|
| 18 |
+
"layer": 2,
|
| 19 |
+
"refusal_score": 8.61406135559082,
|
| 20 |
+
"steering_score": -11.801252365112305,
|
| 21 |
+
"kl_div_score": 0.02662166039541102
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"position": -5,
|
| 25 |
+
"layer": 3,
|
| 26 |
+
"refusal_score": 8.32544994354248,
|
| 27 |
+
"steering_score": -11.922663688659668,
|
| 28 |
+
"kl_div_score": 0.06110219972943868
|
| 29 |
+
},
|
| 30 |
+
{
|
| 31 |
+
"position": -5,
|
| 32 |
+
"layer": 4,
|
| 33 |
+
"refusal_score": 7.7651824951171875,
|
| 34 |
+
"steering_score": -11.739885330200195,
|
| 35 |
+
"kl_div_score": 0.06489012377394329
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"position": -5,
|
| 39 |
+
"layer": 5,
|
| 40 |
+
"refusal_score": 5.078960418701172,
|
| 41 |
+
"steering_score": -11.384549140930176,
|
| 42 |
+
"kl_div_score": 0.055092546867949475
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"position": -5,
|
| 46 |
+
"layer": 6,
|
| 47 |
+
"refusal_score": 4.14361572265625,
|
| 48 |
+
"steering_score": -10.486276626586914,
|
| 49 |
+
"kl_div_score": 0.05534075754782512
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"position": -5,
|
| 53 |
+
"layer": 7,
|
| 54 |
+
"refusal_score": 0.5053978562355042,
|
| 55 |
+
"steering_score": -7.856618881225586,
|
| 56 |
+
"kl_div_score": 0.12158955980096406
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"position": -5,
|
| 60 |
+
"layer": 8,
|
| 61 |
+
"refusal_score": -3.547269344329834,
|
| 62 |
+
"steering_score": -1.4370272159576416,
|
| 63 |
+
"kl_div_score": 0.13569090034394904
|
| 64 |
+
},
|
| 65 |
+
{
|
| 66 |
+
"position": -5,
|
| 67 |
+
"layer": 9,
|
| 68 |
+
"refusal_score": -4.746824264526367,
|
| 69 |
+
"steering_score": 2.9055585861206055,
|
| 70 |
+
"kl_div_score": 0.1485775017862303
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"position": -5,
|
| 74 |
+
"layer": 10,
|
| 75 |
+
"refusal_score": -5.376955032348633,
|
| 76 |
+
"steering_score": 7.028485298156738,
|
| 77 |
+
"kl_div_score": 0.08796039998349871
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"position": -5,
|
| 81 |
+
"layer": 11,
|
| 82 |
+
"refusal_score": -8.539012908935547,
|
| 83 |
+
"steering_score": 8.075133323669434,
|
| 84 |
+
"kl_div_score": 0.06541798089942039
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"position": -5,
|
| 88 |
+
"layer": 12,
|
| 89 |
+
"refusal_score": -10.165912628173828,
|
| 90 |
+
"steering_score": 8.994918823242188,
|
| 91 |
+
"kl_div_score": 0.1132835685413567
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"position": -5,
|
| 95 |
+
"layer": 13,
|
| 96 |
+
"refusal_score": -10.355609893798828,
|
| 97 |
+
"steering_score": 5.637273788452148,
|
| 98 |
+
"kl_div_score": 0.2653944454585543
|
| 99 |
+
},
|
| 100 |
+
{
|
| 101 |
+
"position": -5,
|
| 102 |
+
"layer": 14,
|
| 103 |
+
"refusal_score": -6.636038303375244,
|
| 104 |
+
"steering_score": 4.687836170196533,
|
| 105 |
+
"kl_div_score": 0.0141100183367428
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"position": -5,
|
| 109 |
+
"layer": 15,
|
| 110 |
+
"refusal_score": -5.5487470626831055,
|
| 111 |
+
"steering_score": 2.144282102584839,
|
| 112 |
+
"kl_div_score": 0.02395068876429188
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"position": -5,
|
| 116 |
+
"layer": 16,
|
| 117 |
+
"refusal_score": -5.478135108947754,
|
| 118 |
+
"steering_score": 1.0938483476638794,
|
| 119 |
+
"kl_div_score": 0.012701942605351556
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"position": -5,
|
| 123 |
+
"layer": 17,
|
| 124 |
+
"refusal_score": -5.213899612426758,
|
| 125 |
+
"steering_score": -0.17677417397499084,
|
| 126 |
+
"kl_div_score": 0.013754030688354424
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"position": -5,
|
| 130 |
+
"layer": 18,
|
| 131 |
+
"refusal_score": -6.074069976806641,
|
| 132 |
+
"steering_score": -1.1289219856262207,
|
| 133 |
+
"kl_div_score": 0.011300947735016189
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"position": -5,
|
| 137 |
+
"layer": 19,
|
| 138 |
+
"refusal_score": -1.6875636577606201,
|
| 139 |
+
"steering_score": -2.335386276245117,
|
| 140 |
+
"kl_div_score": 0.024858671554434622
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"position": -5,
|
| 144 |
+
"layer": 20,
|
| 145 |
+
"refusal_score": -2.975120782852173,
|
| 146 |
+
"steering_score": -2.684943914413452,
|
| 147 |
+
"kl_div_score": 0.012842250835919534
|
| 148 |
+
},
|
| 149 |
+
{
|
| 150 |
+
"position": -5,
|
| 151 |
+
"layer": 21,
|
| 152 |
+
"refusal_score": -3.183622360229492,
|
| 153 |
+
"steering_score": -3.214367628097534,
|
| 154 |
+
"kl_div_score": 0.021878747008857467
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"position": -5,
|
| 158 |
+
"layer": 22,
|
| 159 |
+
"refusal_score": -1.722489595413208,
|
| 160 |
+
"steering_score": -3.474200487136841,
|
| 161 |
+
"kl_div_score": 0.022382147956212104
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"position": -5,
|
| 165 |
+
"layer": 23,
|
| 166 |
+
"refusal_score": -0.6202753186225891,
|
| 167 |
+
"steering_score": -4.342594146728516,
|
| 168 |
+
"kl_div_score": 0.017418699702721193
|
| 169 |
+
},
|
| 170 |
+
{
|
| 171 |
+
"position": -5,
|
| 172 |
+
"layer": 24,
|
| 173 |
+
"refusal_score": 1.2065739631652832,
|
| 174 |
+
"steering_score": -4.2968645095825195,
|
| 175 |
+
"kl_div_score": 0.016258465730455446
|
| 176 |
+
},
|
| 177 |
+
{
|
| 178 |
+
"position": -5,
|
| 179 |
+
"layer": 25,
|
| 180 |
+
"refusal_score": 1.9451597929000854,
|
| 181 |
+
"steering_score": -4.316882133483887,
|
| 182 |
+
"kl_div_score": 0.013315564529131926
|
| 183 |
+
},
|
| 184 |
+
{
|
| 185 |
+
"position": -5,
|
| 186 |
+
"layer": 26,
|
| 187 |
+
"refusal_score": 2.7124099731445312,
|
| 188 |
+
"steering_score": -4.45754861831665,
|
| 189 |
+
"kl_div_score": 0.011776859122535012
|
| 190 |
+
},
|
| 191 |
+
{
|
| 192 |
+
"position": -5,
|
| 193 |
+
"layer": 27,
|
| 194 |
+
"refusal_score": 4.2641096115112305,
|
| 195 |
+
"steering_score": -4.595405578613281,
|
| 196 |
+
"kl_div_score": 0.009685238960157085
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"position": -5,
|
| 200 |
+
"layer": 28,
|
| 201 |
+
"refusal_score": 5.035680770874023,
|
| 202 |
+
"steering_score": -4.640095233917236,
|
| 203 |
+
"kl_div_score": 0.011553904695468277
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"position": -5,
|
| 207 |
+
"layer": 29,
|
| 208 |
+
"refusal_score": 4.5578694343566895,
|
| 209 |
+
"steering_score": -4.819130897521973,
|
| 210 |
+
"kl_div_score": 0.009942203119875617
|
| 211 |
+
},
|
| 212 |
+
{
|
| 213 |
+
"position": -5,
|
| 214 |
+
"layer": 30,
|
| 215 |
+
"refusal_score": 4.937906265258789,
|
| 216 |
+
"steering_score": -4.4534406661987305,
|
| 217 |
+
"kl_div_score": 0.010485377575664055
|
| 218 |
+
},
|
| 219 |
+
{
|
| 220 |
+
"position": -5,
|
| 221 |
+
"layer": 31,
|
| 222 |
+
"refusal_score": 5.580471038818359,
|
| 223 |
+
"steering_score": -5.183733940124512,
|
| 224 |
+
"kl_div_score": 0.0073827847249879185
|
| 225 |
+
},
|
| 226 |
+
{
|
| 227 |
+
"position": -4,
|
| 228 |
+
"layer": 0,
|
| 229 |
+
"refusal_score": 7.450636863708496,
|
| 230 |
+
"steering_score": -11.830456733703613,
|
| 231 |
+
"kl_div_score": 0.0
|
| 232 |
+
},
|
| 233 |
+
{
|
| 234 |
+
"position": -4,
|
| 235 |
+
"layer": 1,
|
| 236 |
+
"refusal_score": 10.812314987182617,
|
| 237 |
+
"steering_score": -11.795536994934082,
|
| 238 |
+
"kl_div_score": 0.4158383283900479
|
| 239 |
+
},
|
| 240 |
+
{
|
| 241 |
+
"position": -4,
|
| 242 |
+
"layer": 2,
|
| 243 |
+
"refusal_score": 6.827181816101074,
|
| 244 |
+
"steering_score": -11.737068176269531,
|
| 245 |
+
"kl_div_score": 0.20311702320501046
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"position": -4,
|
| 249 |
+
"layer": 3,
|
| 250 |
+
"refusal_score": 8.025232315063477,
|
| 251 |
+
"steering_score": -11.748991966247559,
|
| 252 |
+
"kl_div_score": 0.02712648312854689
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"position": -4,
|
| 256 |
+
"layer": 4,
|
| 257 |
+
"refusal_score": 6.425276279449463,
|
| 258 |
+
"steering_score": -11.467093467712402,
|
| 259 |
+
"kl_div_score": 0.006255747482580087
|
| 260 |
+
},
|
| 261 |
+
{
|
| 262 |
+
"position": -4,
|
| 263 |
+
"layer": 5,
|
| 264 |
+
"refusal_score": 6.909249305725098,
|
| 265 |
+
"steering_score": -10.97586441040039,
|
| 266 |
+
"kl_div_score": 0.009888444520913529
|
| 267 |
+
},
|
| 268 |
+
{
|
| 269 |
+
"position": -4,
|
| 270 |
+
"layer": 6,
|
| 271 |
+
"refusal_score": 4.926372528076172,
|
| 272 |
+
"steering_score": -10.25048828125,
|
| 273 |
+
"kl_div_score": 0.01978332223716977
|
| 274 |
+
},
|
| 275 |
+
{
|
| 276 |
+
"position": -4,
|
| 277 |
+
"layer": 7,
|
| 278 |
+
"refusal_score": 6.651894569396973,
|
| 279 |
+
"steering_score": -9.241510391235352,
|
| 280 |
+
"kl_div_score": 0.04994251613248886
|
| 281 |
+
},
|
| 282 |
+
{
|
| 283 |
+
"position": -4,
|
| 284 |
+
"layer": 8,
|
| 285 |
+
"refusal_score": 6.573302745819092,
|
| 286 |
+
"steering_score": -6.214093208312988,
|
| 287 |
+
"kl_div_score": 0.042811360122624306
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"position": -4,
|
| 291 |
+
"layer": 9,
|
| 292 |
+
"refusal_score": 6.5962138175964355,
|
| 293 |
+
"steering_score": -3.7241485118865967,
|
| 294 |
+
"kl_div_score": 0.033524626113400284
|
| 295 |
+
},
|
| 296 |
+
{
|
| 297 |
+
"position": -4,
|
| 298 |
+
"layer": 10,
|
| 299 |
+
"refusal_score": 4.0028791427612305,
|
| 300 |
+
"steering_score": 4.6848249435424805,
|
| 301 |
+
"kl_div_score": 0.03246775216125604
|
| 302 |
+
},
|
| 303 |
+
{
|
| 304 |
+
"position": -4,
|
| 305 |
+
"layer": 11,
|
| 306 |
+
"refusal_score": 4.505997657775879,
|
| 307 |
+
"steering_score": 3.1285815238952637,
|
| 308 |
+
"kl_div_score": 0.03856063620648774
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"position": -4,
|
| 312 |
+
"layer": 12,
|
| 313 |
+
"refusal_score": 3.0150375366210938,
|
| 314 |
+
"steering_score": 5.0643134117126465,
|
| 315 |
+
"kl_div_score": 0.04881445734757695
|
| 316 |
+
},
|
| 317 |
+
{
|
| 318 |
+
"position": -4,
|
| 319 |
+
"layer": 13,
|
| 320 |
+
"refusal_score": 1.1747349500656128,
|
| 321 |
+
"steering_score": 1.4116528034210205,
|
| 322 |
+
"kl_div_score": 0.026868538149248188
|
| 323 |
+
},
|
| 324 |
+
{
|
| 325 |
+
"position": -4,
|
| 326 |
+
"layer": 14,
|
| 327 |
+
"refusal_score": 3.023860454559326,
|
| 328 |
+
"steering_score": -3.3504014015197754,
|
| 329 |
+
"kl_div_score": 0.04032019237447497
|
| 330 |
+
},
|
| 331 |
+
{
|
| 332 |
+
"position": -4,
|
| 333 |
+
"layer": 15,
|
| 334 |
+
"refusal_score": 5.741189956665039,
|
| 335 |
+
"steering_score": -5.515679359436035,
|
| 336 |
+
"kl_div_score": 0.12204382576863267
|
| 337 |
+
},
|
| 338 |
+
{
|
| 339 |
+
"position": -4,
|
| 340 |
+
"layer": 16,
|
| 341 |
+
"refusal_score": 4.668534755706787,
|
| 342 |
+
"steering_score": -6.197427272796631,
|
| 343 |
+
"kl_div_score": 0.020614937937280454
|
| 344 |
+
},
|
| 345 |
+
{
|
| 346 |
+
"position": -4,
|
| 347 |
+
"layer": 17,
|
| 348 |
+
"refusal_score": 5.006313323974609,
|
| 349 |
+
"steering_score": -7.727458953857422,
|
| 350 |
+
"kl_div_score": 0.014675440656217304
|
| 351 |
+
},
|
| 352 |
+
{
|
| 353 |
+
"position": -4,
|
| 354 |
+
"layer": 18,
|
| 355 |
+
"refusal_score": 5.147755146026611,
|
| 356 |
+
"steering_score": -7.854608058929443,
|
| 357 |
+
"kl_div_score": 0.013922934527556491
|
| 358 |
+
},
|
| 359 |
+
{
|
| 360 |
+
"position": -4,
|
| 361 |
+
"layer": 19,
|
| 362 |
+
"refusal_score": 5.385741233825684,
|
| 363 |
+
"steering_score": -8.62336540222168,
|
| 364 |
+
"kl_div_score": 0.006841440913400462
|
| 365 |
+
},
|
| 366 |
+
{
|
| 367 |
+
"position": -4,
|
| 368 |
+
"layer": 20,
|
| 369 |
+
"refusal_score": 4.483187675476074,
|
| 370 |
+
"steering_score": -8.518564224243164,
|
| 371 |
+
"kl_div_score": 0.018073221561273382
|
| 372 |
+
},
|
| 373 |
+
{
|
| 374 |
+
"position": -4,
|
| 375 |
+
"layer": 21,
|
| 376 |
+
"refusal_score": 4.432735919952393,
|
| 377 |
+
"steering_score": -8.875178337097168,
|
| 378 |
+
"kl_div_score": 0.03618557665671606
|
| 379 |
+
},
|
| 380 |
+
{
|
| 381 |
+
"position": -4,
|
| 382 |
+
"layer": 22,
|
| 383 |
+
"refusal_score": 3.882089614868164,
|
| 384 |
+
"steering_score": -9.056167602539062,
|
| 385 |
+
"kl_div_score": 0.029761310959065115
|
| 386 |
+
},
|
| 387 |
+
{
|
| 388 |
+
"position": -4,
|
| 389 |
+
"layer": 23,
|
| 390 |
+
"refusal_score": 4.477509498596191,
|
| 391 |
+
"steering_score": -9.458629608154297,
|
| 392 |
+
"kl_div_score": 0.04115770017240697
|
| 393 |
+
},
|
| 394 |
+
{
|
| 395 |
+
"position": -4,
|
| 396 |
+
"layer": 24,
|
| 397 |
+
"refusal_score": 5.272890090942383,
|
| 398 |
+
"steering_score": -9.620115280151367,
|
| 399 |
+
"kl_div_score": 0.026080006169360244
|
| 400 |
+
},
|
| 401 |
+
{
|
| 402 |
+
"position": -4,
|
| 403 |
+
"layer": 25,
|
| 404 |
+
"refusal_score": 5.76324462890625,
|
| 405 |
+
"steering_score": -9.744662284851074,
|
| 406 |
+
"kl_div_score": 0.025975317857587155
|
| 407 |
+
},
|
| 408 |
+
{
|
| 409 |
+
"position": -4,
|
| 410 |
+
"layer": 26,
|
| 411 |
+
"refusal_score": 5.6695556640625,
|
| 412 |
+
"steering_score": -9.72978687286377,
|
| 413 |
+
"kl_div_score": 0.021038816397045645
|
| 414 |
+
},
|
| 415 |
+
{
|
| 416 |
+
"position": -4,
|
| 417 |
+
"layer": 27,
|
| 418 |
+
"refusal_score": 6.13779354095459,
|
| 419 |
+
"steering_score": -9.982513427734375,
|
| 420 |
+
"kl_div_score": 0.01604703955541417
|
| 421 |
+
},
|
| 422 |
+
{
|
| 423 |
+
"position": -4,
|
| 424 |
+
"layer": 28,
|
| 425 |
+
"refusal_score": 6.53122615814209,
|
| 426 |
+
"steering_score": -10.206818580627441,
|
| 427 |
+
"kl_div_score": 0.00905406421169553
|
| 428 |
+
},
|
| 429 |
+
{
|
| 430 |
+
"position": -4,
|
| 431 |
+
"layer": 29,
|
| 432 |
+
"refusal_score": 6.67525053024292,
|
| 433 |
+
"steering_score": -10.416938781738281,
|
| 434 |
+
"kl_div_score": 0.010709122069082595
|
| 435 |
+
},
|
| 436 |
+
{
|
| 437 |
+
"position": -4,
|
| 438 |
+
"layer": 30,
|
| 439 |
+
"refusal_score": 6.565028190612793,
|
| 440 |
+
"steering_score": -10.57150936126709,
|
| 441 |
+
"kl_div_score": 0.013185399142395403
|
| 442 |
+
},
|
| 443 |
+
{
|
| 444 |
+
"position": -4,
|
| 445 |
+
"layer": 31,
|
| 446 |
+
"refusal_score": 7.538625717163086,
|
| 447 |
+
"steering_score": -9.432778358459473,
|
| 448 |
+
"kl_div_score": 0.00659994063938069
|
| 449 |
+
},
|
| 450 |
+
{
|
| 451 |
+
"position": -3,
|
| 452 |
+
"layer": 0,
|
| 453 |
+
"refusal_score": 7.450636863708496,
|
| 454 |
+
"steering_score": -11.830456733703613,
|
| 455 |
+
"kl_div_score": 0.0
|
| 456 |
+
},
|
| 457 |
+
{
|
| 458 |
+
"position": -3,
|
| 459 |
+
"layer": 1,
|
| 460 |
+
"refusal_score": 7.351442337036133,
|
| 461 |
+
"steering_score": -11.879144668579102,
|
| 462 |
+
"kl_div_score": 0.05714835705792163
|
| 463 |
+
},
|
| 464 |
+
{
|
| 465 |
+
"position": -3,
|
| 466 |
+
"layer": 2,
|
| 467 |
+
"refusal_score": 7.555194854736328,
|
| 468 |
+
"steering_score": -11.859851837158203,
|
| 469 |
+
"kl_div_score": 0.018264240284618614
|
| 470 |
+
},
|
| 471 |
+
{
|
| 472 |
+
"position": -3,
|
| 473 |
+
"layer": 3,
|
| 474 |
+
"refusal_score": 7.613396167755127,
|
| 475 |
+
"steering_score": -12.02131462097168,
|
| 476 |
+
"kl_div_score": 0.05679509519728466
|
| 477 |
+
},
|
| 478 |
+
{
|
| 479 |
+
"position": -3,
|
| 480 |
+
"layer": 4,
|
| 481 |
+
"refusal_score": 7.602565765380859,
|
| 482 |
+
"steering_score": -11.988609313964844,
|
| 483 |
+
"kl_div_score": 0.017283959215815804
|
| 484 |
+
},
|
| 485 |
+
{
|
| 486 |
+
"position": -3,
|
| 487 |
+
"layer": 5,
|
| 488 |
+
"refusal_score": 7.565119743347168,
|
| 489 |
+
"steering_score": -12.049741744995117,
|
| 490 |
+
"kl_div_score": 0.023579715899906217
|
| 491 |
+
},
|
| 492 |
+
{
|
| 493 |
+
"position": -3,
|
| 494 |
+
"layer": 6,
|
| 495 |
+
"refusal_score": 7.761242389678955,
|
| 496 |
+
"steering_score": -12.232732772827148,
|
| 497 |
+
"kl_div_score": 0.0022932509288439333
|
| 498 |
+
},
|
| 499 |
+
{
|
| 500 |
+
"position": -3,
|
| 501 |
+
"layer": 7,
|
| 502 |
+
"refusal_score": 7.998626708984375,
|
| 503 |
+
"steering_score": -12.119717597961426,
|
| 504 |
+
"kl_div_score": 0.0026027950945717204
|
| 505 |
+
},
|
| 506 |
+
{
|
| 507 |
+
"position": -3,
|
| 508 |
+
"layer": 8,
|
| 509 |
+
"refusal_score": 7.18452262878418,
|
| 510 |
+
"steering_score": -12.026199340820312,
|
| 511 |
+
"kl_div_score": 0.014272619988313409
|
| 512 |
+
},
|
| 513 |
+
{
|
| 514 |
+
"position": -3,
|
| 515 |
+
"layer": 9,
|
| 516 |
+
"refusal_score": 6.099905967712402,
|
| 517 |
+
"steering_score": -11.814745903015137,
|
| 518 |
+
"kl_div_score": 0.022906798340328747
|
| 519 |
+
},
|
| 520 |
+
{
|
| 521 |
+
"position": -3,
|
| 522 |
+
"layer": 10,
|
| 523 |
+
"refusal_score": 1.6819469928741455,
|
| 524 |
+
"steering_score": -10.3506441116333,
|
| 525 |
+
"kl_div_score": 0.02854894664135504
|
| 526 |
+
},
|
| 527 |
+
{
|
| 528 |
+
"position": -3,
|
| 529 |
+
"layer": 11,
|
| 530 |
+
"refusal_score": -7.098132610321045,
|
| 531 |
+
"steering_score": -1.9208875894546509,
|
| 532 |
+
"kl_div_score": 0.03871730206055337
|
| 533 |
+
},
|
| 534 |
+
{
|
| 535 |
+
"position": -3,
|
| 536 |
+
"layer": 12,
|
| 537 |
+
"refusal_score": -9.373048782348633,
|
| 538 |
+
"steering_score": 3.060147523880005,
|
| 539 |
+
"kl_div_score": 0.0867584559325129
|
| 540 |
+
},
|
| 541 |
+
{
|
| 542 |
+
"position": -3,
|
| 543 |
+
"layer": 13,
|
| 544 |
+
"refusal_score": -5.892275810241699,
|
| 545 |
+
"steering_score": 0.07500703632831573,
|
| 546 |
+
"kl_div_score": 0.1633262984950845
|
| 547 |
+
},
|
| 548 |
+
{
|
| 549 |
+
"position": -3,
|
| 550 |
+
"layer": 14,
|
| 551 |
+
"refusal_score": -0.7610650062561035,
|
| 552 |
+
"steering_score": 2.843015193939209,
|
| 553 |
+
"kl_div_score": 0.01738917049392854
|
| 554 |
+
},
|
| 555 |
+
{
|
| 556 |
+
"position": -3,
|
| 557 |
+
"layer": 15,
|
| 558 |
+
"refusal_score": -0.5506424307823181,
|
| 559 |
+
"steering_score": 3.8185596466064453,
|
| 560 |
+
"kl_div_score": 0.0688840249624754
|
| 561 |
+
},
|
| 562 |
+
{
|
| 563 |
+
"position": -3,
|
| 564 |
+
"layer": 16,
|
| 565 |
+
"refusal_score": 0.20622670650482178,
|
| 566 |
+
"steering_score": 2.1972856521606445,
|
| 567 |
+
"kl_div_score": 0.1475035138357805
|
| 568 |
+
},
|
| 569 |
+
{
|
| 570 |
+
"position": -3,
|
| 571 |
+
"layer": 17,
|
| 572 |
+
"refusal_score": 1.5745298862457275,
|
| 573 |
+
"steering_score": -0.0007034424343146384,
|
| 574 |
+
"kl_div_score": 0.680177021099812
|
| 575 |
+
},
|
| 576 |
+
{
|
| 577 |
+
"position": -3,
|
| 578 |
+
"layer": 18,
|
| 579 |
+
"refusal_score": 2.412114381790161,
|
| 580 |
+
"steering_score": -0.9559683799743652,
|
| 581 |
+
"kl_div_score": 0.6144676889726809
|
| 582 |
+
},
|
| 583 |
+
{
|
| 584 |
+
"position": -3,
|
| 585 |
+
"layer": 19,
|
| 586 |
+
"refusal_score": 3.371020793914795,
|
| 587 |
+
"steering_score": -1.605437159538269,
|
| 588 |
+
"kl_div_score": 1.087222243589543
|
| 589 |
+
},
|
| 590 |
+
{
|
| 591 |
+
"position": -3,
|
| 592 |
+
"layer": 20,
|
| 593 |
+
"refusal_score": 3.380979061126709,
|
| 594 |
+
"steering_score": -1.6221394538879395,
|
| 595 |
+
"kl_div_score": 0.8293832443842403
|
| 596 |
+
},
|
| 597 |
+
{
|
| 598 |
+
"position": -3,
|
| 599 |
+
"layer": 21,
|
| 600 |
+
"refusal_score": 2.6941325664520264,
|
| 601 |
+
"steering_score": -1.701312780380249,
|
| 602 |
+
"kl_div_score": 0.5683925898575501
|
| 603 |
+
},
|
| 604 |
+
{
|
| 605 |
+
"position": -3,
|
| 606 |
+
"layer": 22,
|
| 607 |
+
"refusal_score": 3.3659584522247314,
|
| 608 |
+
"steering_score": -1.4302847385406494,
|
| 609 |
+
"kl_div_score": 0.5506433252182473
|
| 610 |
+
},
|
| 611 |
+
{
|
| 612 |
+
"position": -3,
|
| 613 |
+
"layer": 23,
|
| 614 |
+
"refusal_score": 3.5170602798461914,
|
| 615 |
+
"steering_score": -1.8237537145614624,
|
| 616 |
+
"kl_div_score": 0.5088802263664781
|
| 617 |
+
},
|
| 618 |
+
{
|
| 619 |
+
"position": -3,
|
| 620 |
+
"layer": 24,
|
| 621 |
+
"refusal_score": 3.5988001823425293,
|
| 622 |
+
"steering_score": -1.5716158151626587,
|
| 623 |
+
"kl_div_score": 0.4243232694243198
|
| 624 |
+
},
|
| 625 |
+
{
|
| 626 |
+
"position": -3,
|
| 627 |
+
"layer": 25,
|
| 628 |
+
"refusal_score": 3.224271774291992,
|
| 629 |
+
"steering_score": -1.4275472164154053,
|
| 630 |
+
"kl_div_score": 0.32617104795758256
|
| 631 |
+
},
|
| 632 |
+
{
|
| 633 |
+
"position": -3,
|
| 634 |
+
"layer": 26,
|
| 635 |
+
"refusal_score": 3.0915687084198,
|
| 636 |
+
"steering_score": -1.7615240812301636,
|
| 637 |
+
"kl_div_score": 0.24233564495948606
|
| 638 |
+
},
|
| 639 |
+
{
|
| 640 |
+
"position": -3,
|
| 641 |
+
"layer": 27,
|
| 642 |
+
"refusal_score": 3.930178642272949,
|
| 643 |
+
"steering_score": -2.097964286804199,
|
| 644 |
+
"kl_div_score": 0.2857087263265544
|
| 645 |
+
},
|
| 646 |
+
{
|
| 647 |
+
"position": -3,
|
| 648 |
+
"layer": 28,
|
| 649 |
+
"refusal_score": 4.111324310302734,
|
| 650 |
+
"steering_score": -2.3234660625457764,
|
| 651 |
+
"kl_div_score": 0.17600789216823837
|
| 652 |
+
},
|
| 653 |
+
{
|
| 654 |
+
"position": -3,
|
| 655 |
+
"layer": 29,
|
| 656 |
+
"refusal_score": 4.02263879776001,
|
| 657 |
+
"steering_score": -2.341846466064453,
|
| 658 |
+
"kl_div_score": 0.11356473608828227
|
| 659 |
+
},
|
| 660 |
+
{
|
| 661 |
+
"position": -3,
|
| 662 |
+
"layer": 30,
|
| 663 |
+
"refusal_score": 5.118901252746582,
|
| 664 |
+
"steering_score": -2.601374387741089,
|
| 665 |
+
"kl_div_score": 0.037726283129610017
|
| 666 |
+
},
|
| 667 |
+
{
|
| 668 |
+
"position": -3,
|
| 669 |
+
"layer": 31,
|
| 670 |
+
"refusal_score": 7.2652153968811035,
|
| 671 |
+
"steering_score": -2.683837890625,
|
| 672 |
+
"kl_div_score": 0.012861004017420533
|
| 673 |
+
},
|
| 674 |
+
{
|
| 675 |
+
"position": -2,
|
| 676 |
+
"layer": 0,
|
| 677 |
+
"refusal_score": 7.450636863708496,
|
| 678 |
+
"steering_score": -11.830456733703613,
|
| 679 |
+
"kl_div_score": 0.0
|
| 680 |
+
},
|
| 681 |
+
{
|
| 682 |
+
"position": -2,
|
| 683 |
+
"layer": 1,
|
| 684 |
+
"refusal_score": 6.926996231079102,
|
| 685 |
+
"steering_score": -11.798026084899902,
|
| 686 |
+
"kl_div_score": 0.5867804379978319
|
| 687 |
+
},
|
| 688 |
+
{
|
| 689 |
+
"position": -2,
|
| 690 |
+
"layer": 2,
|
| 691 |
+
"refusal_score": -3.8163418769836426,
|
| 692 |
+
"steering_score": -11.823830604553223,
|
| 693 |
+
"kl_div_score": 1.6575609711285022
|
| 694 |
+
},
|
| 695 |
+
{
|
| 696 |
+
"position": -2,
|
| 697 |
+
"layer": 3,
|
| 698 |
+
"refusal_score": 8.180908203125,
|
| 699 |
+
"steering_score": -11.759471893310547,
|
| 700 |
+
"kl_div_score": 0.04485797158089392
|
| 701 |
+
},
|
| 702 |
+
{
|
| 703 |
+
"position": -2,
|
| 704 |
+
"layer": 4,
|
| 705 |
+
"refusal_score": 8.299246788024902,
|
| 706 |
+
"steering_score": -11.92286205291748,
|
| 707 |
+
"kl_div_score": 0.042189485187598366
|
| 708 |
+
},
|
| 709 |
+
{
|
| 710 |
+
"position": -2,
|
| 711 |
+
"layer": 5,
|
| 712 |
+
"refusal_score": 8.099535942077637,
|
| 713 |
+
"steering_score": -11.920945167541504,
|
| 714 |
+
"kl_div_score": 0.013529918829430384
|
| 715 |
+
},
|
| 716 |
+
{
|
| 717 |
+
"position": -2,
|
| 718 |
+
"layer": 6,
|
| 719 |
+
"refusal_score": 6.200526237487793,
|
| 720 |
+
"steering_score": -11.95439624786377,
|
| 721 |
+
"kl_div_score": 0.043231514354219576
|
| 722 |
+
},
|
| 723 |
+
{
|
| 724 |
+
"position": -2,
|
| 725 |
+
"layer": 7,
|
| 726 |
+
"refusal_score": 2.4409306049346924,
|
| 727 |
+
"steering_score": -12.042317390441895,
|
| 728 |
+
"kl_div_score": 0.07690331220794447
|
| 729 |
+
},
|
| 730 |
+
{
|
| 731 |
+
"position": -2,
|
| 732 |
+
"layer": 8,
|
| 733 |
+
"refusal_score": -0.9726417660713196,
|
| 734 |
+
"steering_score": -11.459531784057617,
|
| 735 |
+
"kl_div_score": 0.07554150489178757
|
| 736 |
+
},
|
| 737 |
+
{
|
| 738 |
+
"position": -2,
|
| 739 |
+
"layer": 9,
|
| 740 |
+
"refusal_score": -6.368747234344482,
|
| 741 |
+
"steering_score": -6.816261291503906,
|
| 742 |
+
"kl_div_score": 0.2651261668661564
|
| 743 |
+
},
|
| 744 |
+
{
|
| 745 |
+
"position": -2,
|
| 746 |
+
"layer": 10,
|
| 747 |
+
"refusal_score": -6.758663177490234,
|
| 748 |
+
"steering_score": 0.9893432855606079,
|
| 749 |
+
"kl_div_score": 0.18809623941974607
|
| 750 |
+
},
|
| 751 |
+
{
|
| 752 |
+
"position": -2,
|
| 753 |
+
"layer": 11,
|
| 754 |
+
"refusal_score": -9.023905754089355,
|
| 755 |
+
"steering_score": 8.601274490356445,
|
| 756 |
+
"kl_div_score": 0.04104435475014729
|
| 757 |
+
},
|
| 758 |
+
{
|
| 759 |
+
"position": -2,
|
| 760 |
+
"layer": 12,
|
| 761 |
+
"refusal_score": -9.75497817993164,
|
| 762 |
+
"steering_score": 6.006311893463135,
|
| 763 |
+
"kl_div_score": 0.031104022506242827
|
| 764 |
+
},
|
| 765 |
+
{
|
| 766 |
+
"position": -2,
|
| 767 |
+
"layer": 13,
|
| 768 |
+
"refusal_score": -0.7074866890907288,
|
| 769 |
+
"steering_score": 2.2404544353485107,
|
| 770 |
+
"kl_div_score": 0.029380356415469547
|
| 771 |
+
},
|
| 772 |
+
{
|
| 773 |
+
"position": -2,
|
| 774 |
+
"layer": 14,
|
| 775 |
+
"refusal_score": 5.199129104614258,
|
| 776 |
+
"steering_score": 3.0659422874450684,
|
| 777 |
+
"kl_div_score": 0.5353163095877882
|
| 778 |
+
},
|
| 779 |
+
{
|
| 780 |
+
"position": -2,
|
| 781 |
+
"layer": 15,
|
| 782 |
+
"refusal_score": 4.9762983322143555,
|
| 783 |
+
"steering_score": 3.254157781600952,
|
| 784 |
+
"kl_div_score": 1.0480167571754937
|
| 785 |
+
},
|
| 786 |
+
{
|
| 787 |
+
"position": -2,
|
| 788 |
+
"layer": 16,
|
| 789 |
+
"refusal_score": 3.6731338500976562,
|
| 790 |
+
"steering_score": 3.8821756839752197,
|
| 791 |
+
"kl_div_score": 1.2368674247797558
|
| 792 |
+
},
|
| 793 |
+
{
|
| 794 |
+
"position": -2,
|
| 795 |
+
"layer": 17,
|
| 796 |
+
"refusal_score": 3.1663296222686768,
|
| 797 |
+
"steering_score": 3.5367674827575684,
|
| 798 |
+
"kl_div_score": 3.093188072529282
|
| 799 |
+
},
|
| 800 |
+
{
|
| 801 |
+
"position": -2,
|
| 802 |
+
"layer": 18,
|
| 803 |
+
"refusal_score": 3.152223587036133,
|
| 804 |
+
"steering_score": 2.5594911575317383,
|
| 805 |
+
"kl_div_score": 1.575859430034067
|
| 806 |
+
},
|
| 807 |
+
{
|
| 808 |
+
"position": -2,
|
| 809 |
+
"layer": 19,
|
| 810 |
+
"refusal_score": 3.236422061920166,
|
| 811 |
+
"steering_score": 1.7847882509231567,
|
| 812 |
+
"kl_div_score": 2.04167027746993
|
| 813 |
+
},
|
| 814 |
+
{
|
| 815 |
+
"position": -2,
|
| 816 |
+
"layer": 20,
|
| 817 |
+
"refusal_score": 2.354421615600586,
|
| 818 |
+
"steering_score": 1.7620494365692139,
|
| 819 |
+
"kl_div_score": 1.7265832841839912
|
| 820 |
+
},
|
| 821 |
+
{
|
| 822 |
+
"position": -2,
|
| 823 |
+
"layer": 21,
|
| 824 |
+
"refusal_score": 0.4912188947200775,
|
| 825 |
+
"steering_score": 2.477295398712158,
|
| 826 |
+
"kl_div_score": 1.3980995152256832
|
| 827 |
+
},
|
| 828 |
+
{
|
| 829 |
+
"position": -2,
|
| 830 |
+
"layer": 22,
|
| 831 |
+
"refusal_score": 0.4995027780532837,
|
| 832 |
+
"steering_score": 3.015876054763794,
|
| 833 |
+
"kl_div_score": 0.8391094951104818
|
| 834 |
+
},
|
| 835 |
+
{
|
| 836 |
+
"position": -2,
|
| 837 |
+
"layer": 23,
|
| 838 |
+
"refusal_score": 0.4436635375022888,
|
| 839 |
+
"steering_score": 1.9265347719192505,
|
| 840 |
+
"kl_div_score": 0.5401840366261479
|
| 841 |
+
},
|
| 842 |
+
{
|
| 843 |
+
"position": -2,
|
| 844 |
+
"layer": 24,
|
| 845 |
+
"refusal_score": 1.1446397304534912,
|
| 846 |
+
"steering_score": 1.4338568449020386,
|
| 847 |
+
"kl_div_score": 0.4640195631718479
|
| 848 |
+
},
|
| 849 |
+
{
|
| 850 |
+
"position": -2,
|
| 851 |
+
"layer": 25,
|
| 852 |
+
"refusal_score": 1.5433824062347412,
|
| 853 |
+
"steering_score": 1.1149771213531494,
|
| 854 |
+
"kl_div_score": 0.3339530314691103
|
| 855 |
+
},
|
| 856 |
+
{
|
| 857 |
+
"position": -2,
|
| 858 |
+
"layer": 26,
|
| 859 |
+
"refusal_score": 1.7812745571136475,
|
| 860 |
+
"steering_score": 0.8563902974128723,
|
| 861 |
+
"kl_div_score": 0.2535115344191417
|
| 862 |
+
},
|
| 863 |
+
{
|
| 864 |
+
"position": -2,
|
| 865 |
+
"layer": 27,
|
| 866 |
+
"refusal_score": 2.4479846954345703,
|
| 867 |
+
"steering_score": 0.3418017029762268,
|
| 868 |
+
"kl_div_score": 0.21052595868841084
|
| 869 |
+
},
|
| 870 |
+
{
|
| 871 |
+
"position": -2,
|
| 872 |
+
"layer": 28,
|
| 873 |
+
"refusal_score": 2.3051071166992188,
|
| 874 |
+
"steering_score": 0.10388462990522385,
|
| 875 |
+
"kl_div_score": 0.15181116784191137
|
| 876 |
+
},
|
| 877 |
+
{
|
| 878 |
+
"position": -2,
|
| 879 |
+
"layer": 29,
|
| 880 |
+
"refusal_score": 1.9919395446777344,
|
| 881 |
+
"steering_score": -0.19721129536628723,
|
| 882 |
+
"kl_div_score": 0.08905450779835926
|
| 883 |
+
},
|
| 884 |
+
{
|
| 885 |
+
"position": -2,
|
| 886 |
+
"layer": 30,
|
| 887 |
+
"refusal_score": 2.893739700317383,
|
| 888 |
+
"steering_score": -0.3327680826187134,
|
| 889 |
+
"kl_div_score": 0.08569501484161374
|
| 890 |
+
},
|
| 891 |
+
{
|
| 892 |
+
"position": -2,
|
| 893 |
+
"layer": 31,
|
| 894 |
+
"refusal_score": 3.7339277267456055,
|
| 895 |
+
"steering_score": -0.45471158623695374,
|
| 896 |
+
"kl_div_score": 0.04856147079387752
|
| 897 |
+
},
|
| 898 |
+
{
|
| 899 |
+
"position": -1,
|
| 900 |
+
"layer": 0,
|
| 901 |
+
"refusal_score": 7.450636863708496,
|
| 902 |
+
"steering_score": -11.830456733703613,
|
| 903 |
+
"kl_div_score": 0.0
|
| 904 |
+
},
|
| 905 |
+
{
|
| 906 |
+
"position": -1,
|
| 907 |
+
"layer": 1,
|
| 908 |
+
"refusal_score": 8.079374313354492,
|
| 909 |
+
"steering_score": -11.79311466217041,
|
| 910 |
+
"kl_div_score": 0.006414912641295136
|
| 911 |
+
},
|
| 912 |
+
{
|
| 913 |
+
"position": -1,
|
| 914 |
+
"layer": 2,
|
| 915 |
+
"refusal_score": 7.303004264831543,
|
| 916 |
+
"steering_score": -11.744900703430176,
|
| 917 |
+
"kl_div_score": 0.031120544594983552
|
| 918 |
+
},
|
| 919 |
+
{
|
| 920 |
+
"position": -1,
|
| 921 |
+
"layer": 3,
|
| 922 |
+
"refusal_score": 7.523357391357422,
|
| 923 |
+
"steering_score": -12.060215950012207,
|
| 924 |
+
"kl_div_score": 0.060775408374946265
|
| 925 |
+
},
|
| 926 |
+
{
|
| 927 |
+
"position": -1,
|
| 928 |
+
"layer": 4,
|
| 929 |
+
"refusal_score": 7.348278999328613,
|
| 930 |
+
"steering_score": -11.976644515991211,
|
| 931 |
+
"kl_div_score": 0.07863737032091514
|
| 932 |
+
},
|
| 933 |
+
{
|
| 934 |
+
"position": -1,
|
| 935 |
+
"layer": 5,
|
| 936 |
+
"refusal_score": 7.338820934295654,
|
| 937 |
+
"steering_score": -11.921005249023438,
|
| 938 |
+
"kl_div_score": 0.012741925993061249
|
| 939 |
+
},
|
| 940 |
+
{
|
| 941 |
+
"position": -1,
|
| 942 |
+
"layer": 6,
|
| 943 |
+
"refusal_score": 6.427075386047363,
|
| 944 |
+
"steering_score": -11.922883987426758,
|
| 945 |
+
"kl_div_score": 0.17200264418275032
|
| 946 |
+
},
|
| 947 |
+
{
|
| 948 |
+
"position": -1,
|
| 949 |
+
"layer": 7,
|
| 950 |
+
"refusal_score": 3.605194568634033,
|
| 951 |
+
"steering_score": -11.695231437683105,
|
| 952 |
+
"kl_div_score": 0.1327529109141617
|
| 953 |
+
},
|
| 954 |
+
{
|
| 955 |
+
"position": -1,
|
| 956 |
+
"layer": 8,
|
| 957 |
+
"refusal_score": -4.327999114990234,
|
| 958 |
+
"steering_score": -8.877398490905762,
|
| 959 |
+
"kl_div_score": 0.2133217491746702
|
| 960 |
+
},
|
| 961 |
+
{
|
| 962 |
+
"position": -1,
|
| 963 |
+
"layer": 9,
|
| 964 |
+
"refusal_score": -6.285901069641113,
|
| 965 |
+
"steering_score": -2.485100269317627,
|
| 966 |
+
"kl_div_score": 0.10089911963005772
|
| 967 |
+
},
|
| 968 |
+
{
|
| 969 |
+
"position": -1,
|
| 970 |
+
"layer": 10,
|
| 971 |
+
"refusal_score": -7.331896781921387,
|
| 972 |
+
"steering_score": 4.227353096008301,
|
| 973 |
+
"kl_div_score": 0.08309234215185003
|
| 974 |
+
},
|
| 975 |
+
{
|
| 976 |
+
"position": -1,
|
| 977 |
+
"layer": 11,
|
| 978 |
+
"refusal_score": -8.494709014892578,
|
| 979 |
+
"steering_score": 7.610245227813721,
|
| 980 |
+
"kl_div_score": 0.014720118389356012
|
| 981 |
+
},
|
| 982 |
+
{
|
| 983 |
+
"position": -1,
|
| 984 |
+
"layer": 12,
|
| 985 |
+
"refusal_score": -8.680732727050781,
|
| 986 |
+
"steering_score": 4.7956414222717285,
|
| 987 |
+
"kl_div_score": 0.01707359135099005
|
| 988 |
+
},
|
| 989 |
+
{
|
| 990 |
+
"position": -1,
|
| 991 |
+
"layer": 13,
|
| 992 |
+
"refusal_score": 0.6532330513000488,
|
| 993 |
+
"steering_score": -0.4470141530036926,
|
| 994 |
+
"kl_div_score": 0.016470555894460795
|
| 995 |
+
},
|
| 996 |
+
{
|
| 997 |
+
"position": -1,
|
| 998 |
+
"layer": 14,
|
| 999 |
+
"refusal_score": 6.329411506652832,
|
| 1000 |
+
"steering_score": 2.30622935295105,
|
| 1001 |
+
"kl_div_score": 0.9479847871943258
|
| 1002 |
+
},
|
| 1003 |
+
{
|
| 1004 |
+
"position": -1,
|
| 1005 |
+
"layer": 15,
|
| 1006 |
+
"refusal_score": 6.0013041496276855,
|
| 1007 |
+
"steering_score": 1.3209009170532227,
|
| 1008 |
+
"kl_div_score": 1.0885456680832901
|
| 1009 |
+
},
|
| 1010 |
+
{
|
| 1011 |
+
"position": -1,
|
| 1012 |
+
"layer": 16,
|
| 1013 |
+
"refusal_score": 4.267302513122559,
|
| 1014 |
+
"steering_score": 2.8033483028411865,
|
| 1015 |
+
"kl_div_score": 0.9073458383306865
|
| 1016 |
+
},
|
| 1017 |
+
{
|
| 1018 |
+
"position": -1,
|
| 1019 |
+
"layer": 17,
|
| 1020 |
+
"refusal_score": 5.26901912689209,
|
| 1021 |
+
"steering_score": 1.9798250198364258,
|
| 1022 |
+
"kl_div_score": 2.3829972756327256
|
| 1023 |
+
},
|
| 1024 |
+
{
|
| 1025 |
+
"position": -1,
|
| 1026 |
+
"layer": 18,
|
| 1027 |
+
"refusal_score": 5.262829780578613,
|
| 1028 |
+
"steering_score": 2.076411008834839,
|
| 1029 |
+
"kl_div_score": 2.364046320178272
|
| 1030 |
+
},
|
| 1031 |
+
{
|
| 1032 |
+
"position": -1,
|
| 1033 |
+
"layer": 19,
|
| 1034 |
+
"refusal_score": 3.3191628456115723,
|
| 1035 |
+
"steering_score": 1.7674345970153809,
|
| 1036 |
+
"kl_div_score": 1.8763310828443507
|
| 1037 |
+
},
|
| 1038 |
+
{
|
| 1039 |
+
"position": -1,
|
| 1040 |
+
"layer": 20,
|
| 1041 |
+
"refusal_score": 2.9167158603668213,
|
| 1042 |
+
"steering_score": 1.4590768814086914,
|
| 1043 |
+
"kl_div_score": 1.6069269699166155
|
| 1044 |
+
},
|
| 1045 |
+
{
|
| 1046 |
+
"position": -1,
|
| 1047 |
+
"layer": 21,
|
| 1048 |
+
"refusal_score": 1.8632879257202148,
|
| 1049 |
+
"steering_score": 2.609332323074341,
|
| 1050 |
+
"kl_div_score": 1.5477564115083355
|
| 1051 |
+
},
|
| 1052 |
+
{
|
| 1053 |
+
"position": -1,
|
| 1054 |
+
"layer": 22,
|
| 1055 |
+
"refusal_score": 2.5213961601257324,
|
| 1056 |
+
"steering_score": 2.937309741973877,
|
| 1057 |
+
"kl_div_score": 1.2337463087795868
|
| 1058 |
+
},
|
| 1059 |
+
{
|
| 1060 |
+
"position": -1,
|
| 1061 |
+
"layer": 23,
|
| 1062 |
+
"refusal_score": 2.130760669708252,
|
| 1063 |
+
"steering_score": 2.94647216796875,
|
| 1064 |
+
"kl_div_score": 1.135543047215232
|
| 1065 |
+
},
|
| 1066 |
+
{
|
| 1067 |
+
"position": -1,
|
| 1068 |
+
"layer": 24,
|
| 1069 |
+
"refusal_score": 1.4255648851394653,
|
| 1070 |
+
"steering_score": 2.6254642009735107,
|
| 1071 |
+
"kl_div_score": 1.0170067442120234
|
| 1072 |
+
},
|
| 1073 |
+
{
|
| 1074 |
+
"position": -1,
|
| 1075 |
+
"layer": 25,
|
| 1076 |
+
"refusal_score": 1.8470709323883057,
|
| 1077 |
+
"steering_score": 2.4709389209747314,
|
| 1078 |
+
"kl_div_score": 0.9048952123050806
|
| 1079 |
+
},
|
| 1080 |
+
{
|
| 1081 |
+
"position": -1,
|
| 1082 |
+
"layer": 26,
|
| 1083 |
+
"refusal_score": 1.4695944786071777,
|
| 1084 |
+
"steering_score": 1.7421600818634033,
|
| 1085 |
+
"kl_div_score": 0.6825789057908486
|
| 1086 |
+
},
|
| 1087 |
+
{
|
| 1088 |
+
"position": -1,
|
| 1089 |
+
"layer": 27,
|
| 1090 |
+
"refusal_score": 1.1728484630584717,
|
| 1091 |
+
"steering_score": 1.8248333930969238,
|
| 1092 |
+
"kl_div_score": 0.49910105954810363
|
| 1093 |
+
},
|
| 1094 |
+
{
|
| 1095 |
+
"position": -1,
|
| 1096 |
+
"layer": 28,
|
| 1097 |
+
"refusal_score": 0.4489290118217468,
|
| 1098 |
+
"steering_score": 1.8028056621551514,
|
| 1099 |
+
"kl_div_score": 0.3759350678814426
|
| 1100 |
+
},
|
| 1101 |
+
{
|
| 1102 |
+
"position": -1,
|
| 1103 |
+
"layer": 29,
|
| 1104 |
+
"refusal_score": -0.6745665073394775,
|
| 1105 |
+
"steering_score": 2.546116828918457,
|
| 1106 |
+
"kl_div_score": 0.22260585876335837
|
| 1107 |
+
},
|
| 1108 |
+
{
|
| 1109 |
+
"position": -1,
|
| 1110 |
+
"layer": 30,
|
| 1111 |
+
"refusal_score": -0.34203040599823,
|
| 1112 |
+
"steering_score": 2.926508665084839,
|
| 1113 |
+
"kl_div_score": 0.20333847684556203
|
| 1114 |
+
},
|
| 1115 |
+
{
|
| 1116 |
+
"position": -1,
|
| 1117 |
+
"layer": 31,
|
| 1118 |
+
"refusal_score": 0.09251364320516586,
|
| 1119 |
+
"steering_score": 2.020570993423462,
|
| 1120 |
+
"kl_div_score": 0.13057973266326595
|
| 1121 |
+
}
|
| 1122 |
+
]
|
select_direction/direction_evaluations_filtered.json
ADDED
|
@@ -0,0 +1,128 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"position": -2,
|
| 4 |
+
"layer": 12,
|
| 5 |
+
"refusal_score": -9.75497817993164,
|
| 6 |
+
"steering_score": 6.006311893463135,
|
| 7 |
+
"kl_div_score": 0.031104022506242827
|
| 8 |
+
},
|
| 9 |
+
{
|
| 10 |
+
"position": -3,
|
| 11 |
+
"layer": 12,
|
| 12 |
+
"refusal_score": -9.373048782348633,
|
| 13 |
+
"steering_score": 3.060147523880005,
|
| 14 |
+
"kl_div_score": 0.0867584559325129
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"position": -2,
|
| 18 |
+
"layer": 11,
|
| 19 |
+
"refusal_score": -9.023905754089355,
|
| 20 |
+
"steering_score": 8.601274490356445,
|
| 21 |
+
"kl_div_score": 0.04104435475014729
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"position": -1,
|
| 25 |
+
"layer": 12,
|
| 26 |
+
"refusal_score": -8.680732727050781,
|
| 27 |
+
"steering_score": 4.7956414222717285,
|
| 28 |
+
"kl_div_score": 0.01707359135099005
|
| 29 |
+
},
|
| 30 |
+
{
|
| 31 |
+
"position": -5,
|
| 32 |
+
"layer": 11,
|
| 33 |
+
"refusal_score": -8.539012908935547,
|
| 34 |
+
"steering_score": 8.075133323669434,
|
| 35 |
+
"kl_div_score": 0.06541798089942039
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"position": -1,
|
| 39 |
+
"layer": 11,
|
| 40 |
+
"refusal_score": -8.494709014892578,
|
| 41 |
+
"steering_score": 7.610245227813721,
|
| 42 |
+
"kl_div_score": 0.014720118389356012
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"position": -1,
|
| 46 |
+
"layer": 10,
|
| 47 |
+
"refusal_score": -7.331896781921387,
|
| 48 |
+
"steering_score": 4.227353096008301,
|
| 49 |
+
"kl_div_score": 0.08309234215185003
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"position": -5,
|
| 53 |
+
"layer": 14,
|
| 54 |
+
"refusal_score": -6.636038303375244,
|
| 55 |
+
"steering_score": 4.687836170196533,
|
| 56 |
+
"kl_div_score": 0.0141100183367428
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"position": -5,
|
| 60 |
+
"layer": 15,
|
| 61 |
+
"refusal_score": -5.5487470626831055,
|
| 62 |
+
"steering_score": 2.144282102584839,
|
| 63 |
+
"kl_div_score": 0.02395068876429188
|
| 64 |
+
},
|
| 65 |
+
{
|
| 66 |
+
"position": -5,
|
| 67 |
+
"layer": 16,
|
| 68 |
+
"refusal_score": -5.478135108947754,
|
| 69 |
+
"steering_score": 1.0938483476638794,
|
| 70 |
+
"kl_div_score": 0.012701942605351556
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"position": -5,
|
| 74 |
+
"layer": 10,
|
| 75 |
+
"refusal_score": -5.376955032348633,
|
| 76 |
+
"steering_score": 7.028485298156738,
|
| 77 |
+
"kl_div_score": 0.08796039998349871
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"position": -3,
|
| 81 |
+
"layer": 14,
|
| 82 |
+
"refusal_score": -0.7610650062561035,
|
| 83 |
+
"steering_score": 2.843015193939209,
|
| 84 |
+
"kl_div_score": 0.01738917049392854
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"position": -2,
|
| 88 |
+
"layer": 13,
|
| 89 |
+
"refusal_score": -0.7074866890907288,
|
| 90 |
+
"steering_score": 2.2404544353485107,
|
| 91 |
+
"kl_div_score": 0.029380356415469547
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"position": -3,
|
| 95 |
+
"layer": 15,
|
| 96 |
+
"refusal_score": -0.5506424307823181,
|
| 97 |
+
"steering_score": 3.8185596466064453,
|
| 98 |
+
"kl_div_score": 0.0688840249624754
|
| 99 |
+
},
|
| 100 |
+
{
|
| 101 |
+
"position": -4,
|
| 102 |
+
"layer": 13,
|
| 103 |
+
"refusal_score": 1.1747349500656128,
|
| 104 |
+
"steering_score": 1.4116528034210205,
|
| 105 |
+
"kl_div_score": 0.026868538149248188
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"position": -4,
|
| 109 |
+
"layer": 12,
|
| 110 |
+
"refusal_score": 3.0150375366210938,
|
| 111 |
+
"steering_score": 5.0643134117126465,
|
| 112 |
+
"kl_div_score": 0.04881445734757695
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"position": -4,
|
| 116 |
+
"layer": 10,
|
| 117 |
+
"refusal_score": 4.0028791427612305,
|
| 118 |
+
"steering_score": 4.6848249435424805,
|
| 119 |
+
"kl_div_score": 0.03246775216125604
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"position": -4,
|
| 123 |
+
"layer": 11,
|
| 124 |
+
"refusal_score": 4.505997657775879,
|
| 125 |
+
"steering_score": 3.1285815238952637,
|
| 126 |
+
"kl_div_score": 0.03856063620648774
|
| 127 |
+
}
|
| 128 |
+
]
|
select_direction/kl_div_scores.png
ADDED
|