JBShield/interpret.py at main · SeeweedBrain/JBShield · GitHub

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
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
import argparse
import json
from tqdm import tqdm

import nltk
# Download when running for the first time
nltk.download('words')

from config import path_harmful, path_harmless, model_paths
from utils import load_model, load_ori_prompts, get_jailbreak_prompts
from utils import get_sentence_embeddings
from utils import interpret_difference_matrix


def interpret(model_name):
    # Load model
    model, tokenizer = load_model(model_name, model_paths)

    # Load data
    harmful_prompts, harmless_prompts = load_ori_prompts(
        path_harmful, path_harmless
    )
    print("Number of harmful prompts: {}".format(len(harmful_prompts)))
    print("Number of harmless prompts: {}".format(len(harmless_prompts)))
    jailbreaks = [
        "gcg",
        "autodan",
        "saa",
        "drattack",
        "pair",
        "puzzler",
        "ijp",
        "base64",
        "zulu",
    ]
    jailbreak_prompts = get_jailbreak_prompts(model_name, jailbreaks, split="all")
    for jailbreak in jailbreaks:
        print(f"Number of {jailbreak} prompts: {len(jailbreak_prompts[jailbreak])}")

    # Get embdddings for prompts
    harmful_embeddings = get_sentence_embeddings(
        harmful_prompts, model, model_name, tokenizer
    )
    harmless_embeddings = get_sentence_embeddings(
        harmless_prompts, model, model_name, tokenizer
    )

    embeddings_gcg = get_sentence_embeddings(
        jailbreak_prompts["gcg"], model, model_name, tokenizer
    )
    embeddings_puzzler = get_sentence_embeddings(
        jailbreak_prompts["puzzler"], model, model_name, tokenizer
    )
    embeddings_saa = get_sentence_embeddings(
        jailbreak_prompts["saa"], model, model_name, tokenizer
    )
    embeddings_autodan = get_sentence_embeddings(
        jailbreak_prompts["autodan"], model, model_name, tokenizer
    )
    embeddings_drattack = get_sentence_embeddings(
        jailbreak_prompts["drattack"], model, model_name, tokenizer
    )
    embeddings_pair = get_sentence_embeddings(
        jailbreak_prompts["pair"], model, model_name, tokenizer
    )
    embeddings_ijp = get_sentence_embeddings(
        jailbreak_prompts["ijp"], model, model_name, tokenizer
    )
    embeddings_base64 = get_sentence_embeddings(
        jailbreak_prompts["base64"], model, model_name, tokenizer
    )
    embeddings_zulu = get_sentence_embeddings(
        jailbreak_prompts["zulu"], model, model_name, tokenizer
    )

    # Interpret the embeddings
    # Save the results in interpre_results/{model_name}.txt
    with open("interpret_results/{}.txt".format(model_name), "a") as f:
        for layer_n in tqdm(range(len(harmful_embeddings))):
            f.write(
                "====================layer {}====================\n".format(layer_n)
            )
            interpret_tokens, _, _ = interpret_difference_matrix(
                model,
                tokenizer,
                harmful_embeddings[layer_n],
                harmless_embeddings[layer_n],
            )
            f.write("Toxic harmful-harmless " + json.dumps(interpret_tokens) + "\n")
            f.write("------------------------------------------------\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model, tokenizer, embeddings_gcg[layer_n], harmless_embeddings[layer_n]
            )
            f.write("Toxic gcg-harmless " + json.dumps(interpret_tokens) + "\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model, tokenizer, embeddings_saa[layer_n], harmless_embeddings[layer_n]
            )
            f.write("Toxic saa-harmless " + json.dumps(interpret_tokens) + "\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model,
                tokenizer,
                embeddings_autodan[layer_n],
                harmless_embeddings[layer_n],
            )
            f.write("Toxic autodan-harmless " + json.dumps(interpret_tokens) + "\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model,
                tokenizer,
                embeddings_drattack[layer_n],
                harmless_embeddings[layer_n],
            )
            f.write("Toxic drattack-harmless " + json.dumps(interpret_tokens) + "\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model, tokenizer, embeddings_pair[layer_n], harmless_embeddings[layer_n]
            )
            f.write("Toxic pair-harmless " + json.dumps(interpret_tokens) + "\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model,
                tokenizer,
                embeddings_puzzler[layer_n],
                harmless_embeddings[layer_n],
            )
            f.write("Toxic puzzler-harmless " + json.dumps(interpret_tokens) + "\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model, tokenizer, embeddings_ijp[layer_n], harmless_embeddings[layer_n]
            )
            f.write("Toxic ijp-harmless " + json.dumps(interpret_tokens) + "\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model,
                tokenizer,
                embeddings_base64[layer_n],
                harmless_embeddings[layer_n],
            )
            f.write("Toxic base64-harmless " + json.dumps(interpret_tokens) + "\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model, tokenizer, embeddings_zulu[layer_n], harmless_embeddings[layer_n]
            )
            f.write("Toxic zulu-harmless " + json.dumps(interpret_tokens) + "\n")
            f.write("------------------------------------------------\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model, tokenizer, embeddings_gcg[layer_n], harmful_embeddings[layer_n]
            )
            f.write("Jailbreak gcg-harmful " + json.dumps(interpret_tokens) + "\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model, tokenizer, embeddings_saa[layer_n], harmful_embeddings[layer_n]
            )
            f.write("Jailbreak saa-harmful " + json.dumps(interpret_tokens) + "\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model,
                tokenizer,
                embeddings_autodan[layer_n],
                harmful_embeddings[layer_n],
            )
            f.write("Jailbreak autodan-harmful " + json.dumps(interpret_tokens) + "\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model,
                tokenizer,
                embeddings_drattack[layer_n],
                harmful_embeddings[layer_n],
            )
            f.write("Jailbreak drattack-harmful " + json.dumps(interpret_tokens) + "\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model, tokenizer, embeddings_pair[layer_n], harmful_embeddings[layer_n]
            )
            f.write("Jailbreak pair-harmful " + json.dumps(interpret_tokens) + "\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model,
                tokenizer,
                embeddings_puzzler[layer_n],
                harmful_embeddings[layer_n],
            )
            f.write("Jailbreak puzzler-harmful " + json.dumps(interpret_tokens) + "\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model, tokenizer, embeddings_ijp[layer_n], harmful_embeddings[layer_n]
            )
            f.write("Jailbreak ijp-harmful " + json.dumps(interpret_tokens) + "\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model,
                tokenizer,
                embeddings_base64[layer_n],
                harmful_embeddings[layer_n],
            )
            f.write("Jailbreak base64-harmful " + json.dumps(interpret_tokens) + "\n")
            interpret_tokens, _, _ = interpret_difference_matrix(
                model, tokenizer, embeddings_zulu[layer_n], harmful_embeddings[layer_n]
            )
            f.write("Jailbreak zulu-harmful " + json.dumps(interpret_tokens) + "\n")
            f.flush()


if __name__ == "__main__":
    # Get parameters
    parser = argparse.ArgumentParser(description="Interpret the model")
    parser.add_argument("--model", type=str, help="Taregt model")

    args = parser.parse_args()
    model_name = args.model

    # Show the tokens that are most assiciated with the toxic and jailbreak concepts
    # The results are saved in ./interpre_results
    interpret(model_name)
    print("{} Interpretation done.".format(model_name))

# An example for run this script on llama-2
# python interpret.py --model llama-2