continuous-adv-ICL/src/evaluate.py at main · kaustpradalab/continuous-adv-ICL · 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
import argparse, os, torch, numpy as np, json
# import pandas as pd

import utils, attacks, adv_trainers


def get_args():
    parser = argparse.ArgumentParser()

    utils.add_shared_args(parser)
    utils.add_eval_args(parser)

    return parser.parse_args()


def build_alpacaeval(args, logger):
    default_device = "cuda:0"

    lora_cfg = None
    if args.lora_cfg_path is not None:
        lora_cfg = utils.load_yaml(args.lora_cfg_path, f"{args.save_dir}/{args.save_name}_lora-cfg.yaml")
    # model, tokenizer = utils.get_model(args.model_id, default_device, lora_cfg)
    model, tokenizer = utils.get_model(args.model_id, lora_cfg=lora_cfg)
    if args.model_resume_path is not None:
        logger.info(f"resume model from {args.model_resume_path}")
        state_dict = utils.load_state_dict(args.model_resume_path, mode="safetensors")
        model.load_state_dict(state_dict)
        del state_dict

    alpacaeval_cfg = utils.load_yaml(args.alpacaeval_cfg_path)

    if "datacollator_config" not in alpacaeval_cfg.keys():
        alpacaeval_cfg["datacollator_config"] = {}
    alpacaeval_cfg["datacollator_config"]["output_raw"] = True

    if args.datacollator is not None:
        alpacaeval_cfg["datacollator_config"]["name"] = args.datacollator

    utils.save_yaml(alpacaeval_cfg, f"{args.save_dir}/{args.save_name}_alpacaeval-cfg.yaml")

    alpacaeval_cfg = adv_trainers.AlpacaEvalConfig(**alpacaeval_cfg)

    alpacaeval_set = adv_trainers.build_alpacaeval(
        eval_config=alpacaeval_cfg,
        model=model,
        tokenizer=tokenizer,
        device=default_device,
    )

    json_path = os.path.join(f"{args.save_dir}/{args.save_name}_alpacaeval.json")
    logger.info(f"saving alpacaeval-output to `{json_path}`")
    with open(json_path, "w") as f:
        json.dump(alpacaeval_set, f, indent=2)


def build_evalset(args, logger):
    default_device = "cuda:0"

    lora_cfg = None
    if args.lora_cfg_path is not None:
        lora_cfg = utils.load_yaml(args.lora_cfg_path, f"{args.save_dir}/{args.save_name}_lora-cfg.yaml")
    # model, tokenizer = utils.get_model(args.model_id, default_device, lora_cfg)
    model, tokenizer = utils.get_model(args.model_id, lora_cfg=lora_cfg)
    if args.model_resume_path is not None:
        logger.info(f"resume model from {args.model_resume_path}")
        state_dict = utils.load_state_dict(args.model_resume_path, mode="safetensors")
        model.load_state_dict(state_dict)
        del state_dict

    dataset = utils.get_dataset(args.dataset)

    evalset_cfg = utils.load_yaml(args.evalset_cfg_path)

    # ======== start to override evalset_cfg

    if "datacollator_config" not in evalset_cfg.keys():
        evalset_cfg["datacollator_config"] = {}
    evalset_cfg["datacollator_config"]["output_raw"] = True

    if args.datacollator is not None:
        evalset_cfg["datacollator_config"]["name"] = args.datacollator

    atker_cfg = utils.load_yaml(args.atker_cfg_path)
    if atker_cfg is not None:
        evalset_cfg["attacker_config"] = atker_cfg

    # ======== finished overriding evalset_cfg

    utils.save_yaml(evalset_cfg, f"{args.save_dir}/{args.save_name}_evalset-cfg.yaml")

    evalset_cfg = adv_trainers.EvalsetConfig(**evalset_cfg)

    evalset = adv_trainers.build_evalset(
        eval_config=evalset_cfg,
        model=model,
        tokenizer=tokenizer,
        device=default_device,
        dataset=dataset,
    )

    # df = pd.DataFrame(evalset)

    # csv_path = os.path.join(f"{args.save_dir}/{args.save_name}_evalset.csv")
    # logger.info(f"saving evalset to `{csv_path}`")
    # df.to_csv(csv_path, index=False)

    json_path = os.path.join(f"{args.save_dir}/{args.save_name}_evalset.json")
    logger.info(f"saving adv-evalset to `{json_path}`")
    with open(json_path, "w") as f:
        json.dump(evalset, f, indent=2)


def judge_evalset(args, logger):
    default_device = "cuda:0"

    judger_cfg = utils.load_yaml(args.judger_cfg_path, f"{args.save_dir}/{args.save_name}_judger-cfg.yaml")
    judger_cfg["device"] = default_device
    judger_cfg = adv_trainers.JudgerConfig(**judger_cfg)
    judger = adv_trainers.build_judger(judger_cfg)

    # evalset = pd.read_csv(args.evalset_path)
    with open(args.evalset_path, "r") as f:
        evalset = json.load(f)
    logger.info(f"loaded evalset from {args.evalset_path}")

    # evalset = [
    #     {"prompt": prompt, "generation": generation}
    #     for prompt, generation in zip(evalset["prompt"].tolist(), evalset["generation"].tolist())
    # ]

    # attack_success_rate = judger.judge(evalset)
    # logger.info("attack success rate: {:.3%}".format(attack_success_rate))

    asrs = judger.judge(evalset)
    for k, v in asrs.items():
        logger.info("{}: {:.3%}".format(k, v))


def main(args, logger):
    if args.exp_type == "build-alpacaeval":
        build_alpacaeval(args, logger)
    elif args.exp_type == "build-evalset":
        build_evalset(args, logger)
    elif args.exp_type == "judge-evalset":
        judge_evalset(args, logger)
    else:
        raise ValueError(f"unknown exp-type `{args.exp_type}`")


if __name__ == "__main__":
    args = get_args()
    logger = utils.generic_init(args)

    if args.random_seed is not None:
        np.random.seed(args.random_seed)
        torch.manual_seed(args.random_seed)

    try:
        main(args, logger)
    except Exception as e:
        logger.exception('Unexpected exception! %s', e)