File size: 3,545 Bytes
aefc9ef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env bash
# Copyright (c) Guangsheng Bao.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

# setup the environment
echo `date`, Setup the environment ...
set -e  # exit if error

# prepare folders
exp_path=exp_main_ext
data_path=$exp_path/data
res_path=$exp_path/results
mkdir -p $exp_path $data_path $res_path

datasets="xsum squad writing"
source_models="bloom-7b1 opt-13b llama-13b llama2-13b"

# preparing dataset
for D in $datasets; do
  for M in $source_models; do
    echo `date`, Preparing dataset ${D}_${M} ...
    python scripts/data_builder.py --dataset $D --n_samples 500 --base_model_name $M --output_file $data_path/${D}_${M}
  done
done
exit

# White-box Setting
echo `date`, Evaluate models in the white-box setting:

# evaluate Fast-DetectGPT and fast baselines
for D in $datasets; do
  for M in $source_models; do
    echo `date`, Evaluating Fast-DetectGPT on ${D}_${M} ...
    python scripts/fast_detect_gpt.py --reference_model_name $M --scoring_model_name $M --dataset $D \
                          --dataset_file $data_path/${D}_${M} --output_file $res_path/${D}_${M}

    echo `date`, Evaluating baseline methods on ${D}_${M} ...
    python scripts/baselines.py --scoring_model_name $M --dataset $D \
                          --dataset_file $data_path/${D}_${M} --output_file $res_path/${D}_${M}
  done
done

# evaluate DetectGPT and its improvement DetectLLM
for D in $datasets; do
  for M in $source_models; do
    echo `date`, Evaluating DetectGPT on ${D}_${M} ...
    python scripts/detect_gpt.py --scoring_model_name $M --mask_filling_model_name t5-3b --n_perturbations 100 --dataset $D \
                          --dataset_file $data_path/${D}_${M} --output_file $res_path/${D}_${M}
     # we leverage DetectGPT to generate the perturbations
    echo `date`, Evaluating DetectLLM methods on ${D}_${M} ...
    python scripts/detect_llm.py --scoring_model_name $M --dataset $D \
                          --dataset_file $data_path/${D}_${M}.t5-3b.perturbation_100 --output_file $res_path/${D}_${M}
  done
done


# Black-box Setting
echo `date`, Evaluate models in the black-box setting:
scoring_models="gpt-neo-2.7B"

# evaluate Fast-DetectGPT
for D in $datasets; do
  for M in $source_models; do
    M1=gpt-j-6B  # sampling model
    for M2 in $scoring_models; do
      echo `date`, Evaluating Fast-DetectGPT on ${D}_${M}.${M1}_${M2} ...
      python scripts/fast_detect_gpt.py --reference_model_name ${M1} --scoring_model_name ${M2} --dataset $D \
                          --dataset_file $data_path/${D}_${M} --output_file $res_path/${D}_${M}.${M1}_${M2}
    done
  done
done

# evaluate DetectGPT and its improvement DetectLLM
for D in $datasets; do
  for M in $source_models; do
    M1=t5-3b  # perturbation model
    for M2 in $scoring_models; do
      echo `date`, Evaluating DetectGPT on ${D}_${M}.${M1}_${M2} ...
      python scripts/detect_gpt.py --mask_filling_model_name ${M1} --scoring_model_name ${M2} --n_perturbations 100 --dataset $D \
                          --dataset_file $data_path/${D}_${M} --output_file $res_path/${D}_${M}.${M1}_${M2}
      # we leverage DetectGPT to generate the perturbations
      echo `date`, Evaluating DetectLLM methods on ${D}_${M}.${M1}_${M2} ...
      python scripts/detect_llm.py --scoring_model_name ${M2} --dataset $D \
                          --dataset_file $data_path/${D}_${M}.${M1}.perturbation_100 --output_file $res_path/${D}_${M}.${M1}_${M2}
    done
  done
done