├── .gitignore ├── LICENSE ├── README.md ├── mink_prob.png ├── process_data.py └── src ├── __pycache__ ├── eval.cpython-39.pyc └── options.cpython-39.pyc ├── eval.py ├── options.py ├── out ├── huggyllama │ └── llama-13b_huggyllama │ │ └── llama-7b │ │ └── input │ │ ├── auc.png │ │ ├── auc.txt │ │ ├── auc_google.txt │ │ └── low_google.txt └── text-davinci-003_huggyllama │ └── llama-7b │ └── input │ ├── auc.png │ └── auc.txt └── run.py /.gitignore: -------------------------------------------------------------------------------- 1 | ### AL ### 2 | #Template for AL projects for Dynamics 365 Business Central 3 | #launch.json folder 4 | .vscode/ 5 | #Cache folder 6 | .alcache/ 7 | #Symbols folder 8 | .alpackages/ 9 | #Snapshots folder 10 | .snapshots/ 11 | #Testing Output folder 12 | .output/ 13 | #Extension App-file 14 | *.app 15 | #Rapid Application Development File 16 | rad.json 17 | #Translation Base-file 18 | *.g.xlf 19 | #License-file 20 | *.flf 21 | #Test results file 22 | TestResults.xml -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [yyyy] [name of copyright owner] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # :detective: Detecting Pretraining Data from Large Language Models 2 | 3 | This repository provides an original implementation of [Detecting Pretraining Data from Large Language Models](https://arxiv.org/pdf/2310.16789.pdf) by *Weijia Shi, *Anirudh Ajith, Mengzhou Xia, Yangsibo Huang, Daogao Liu 4 | , Terra Blevins 5 | , Danqi Chen 6 | , Luke Zettlemoyer 7 | 8 | [Website](https://swj0419.github.io/detect-pretrain.github.io/) | [Paper](https://arxiv.org/pdf/2310.16789.pdf) | [WikiMIA Benchmark](https://huggingface.co/datasets/swj0419/WikiMIA) | [BookMIA Benchmark](https://huggingface.co/datasets/swj0419/BookMIA) | [Detection Method Min-K% Prob](#🚀run-our-min-k%-prob-&-other-baselines)(see the following codebase) 9 | 10 | ## Overview 11 | We explore the **pretraining data detection problem**: given a piece of text and black-box access to an LLM without knowing the pretraining data, can we determine if the model was trained on the provided text? 12 | To faciliate the study, we built a dynamic benchmark **WikiMIA** to systematically evaluate detecting methods and proposed **Min-K% Prob** 🕵️, a method for detecting undisclosed pretraining data from large language models. 13 | 14 |

15 | 16 |

17 | 18 | :star: If you find our implementation and paper helpful, please consider citing our work :star: : 19 | 20 | ```bibtex 21 | @misc{shi2023detecting, 22 | title={Detecting Pretraining Data from Large Language Models}, 23 | author={Weijia Shi and Anirudh Ajith and Mengzhou Xia and Yangsibo Huang and Daogao Liu and Terra Blevins and Danqi Chen and Luke Zettlemoyer}, 24 | year={2023}, 25 | eprint={2310.16789}, 26 | archivePrefix={arXiv}, 27 | primaryClass={cs.CL} 28 | } 29 | ``` 30 | 31 | ## 📘 WikiMIA Datasets 32 | 33 | The **WikiMIA datasets** serve as a benchmark designed to evaluate membership inference attack (MIA) methods, specifically in detecting pretraining data from extensive large language models. Access our **WikiMIA datasets** directly on [Hugging Face](https://huggingface.co/datasets/swj0419/WikiMIA). 34 | 35 | #### Loading the Datasets: 36 | 37 | ```python 38 | from datasets import load_dataset 39 | LENGTH = 64 40 | dataset = load_dataset("swj0419/WikiMIA", split=f"WikiMIA_length{LENGTH}") 41 | ``` 42 | * Available Text Lengths: `32, 64, 128, 256`. 43 | * *Label 0*: Refers to the unseen data during pretraining. *Label 1*: Refers to the seen data. 44 | * WikiMIA is applicable to all models released between 2017 to 2023 such as `LLaMA1/2, GPT-Neo, OPT, Pythia, text-davinci-001, text-davinci-002 ...` 45 | 46 | ## 📘 BookMIA Datasets for evaluating MIA on OpenAI models 47 | 48 | The BookMIA datasets serve as a benchmark designed to evaluate membership inference attack (MIA) methods, specifically in detecting pretraining data from OpenAI models that are released before 2023 (such as text-davinci-003). Access our **BookMIA datasets** directly on [Hugging Face](https://huggingface.co/datasets/swj0419/BookMIA). 49 | 50 | The dataset contains non-member and member data: 51 | - non-member data consists of text excerpts from books first published in 2023 52 | - member data includes text excerpts from older books, as categorized by Chang et al. in 2023. 53 | 54 | 55 | #### Loading the Datasets: 56 | 57 | ```python 58 | from datasets import load_dataset 59 | dataset = load_dataset("swj0419/BookMIA") 60 | ``` 61 | * Available Text Lengths: `512`. 62 | * *Label 0*: Refers to the unseen data during pretraining. *Label 1*: Refers to the seen data. 63 | * WikiMIA is applicable to OpenAI models that are released before 2023 `text-davinci-003, text-davinci-002 ...` 64 | 65 | 66 | 67 | ## 🚀 Run our Min-K% Prob & Other Baselines 68 | 69 | Our codebase supports many models: Whether you're using **OpenAI models** that offer logits or models from **Huggingface**, we've got you covered: 70 | 71 | - **OpenAI Models**: 72 | - `text-davinci-003` 73 | - `text-davinci-002` 74 | - ... 75 | 76 | - **Huggingface Models**: 77 | - `meta-llama/Llama-2-70b` 78 | - `huggyllama/llama-70b` 79 | - `EleutherAI/gpt-neox-20b` 80 | - ... 81 | 82 | 🔐 **Important**: When using OpenAI models, ensure to add your API key at `Line 38` in `run.py`: 83 | ```python 84 | openai.api_key = "YOUR_API_KEY" 85 | ``` 86 | Use the following command to run the model: 87 | ```bash 88 | python src/run.py --target_model text-davinci-003 --ref_model huggyllama/llama-7b --data swj0419/WikiMIA --length 64 89 | ``` 90 | 🔍 Parameters Explained: 91 | * Target Model: Set using --target_model. For instance, --target_model huggyllama/llama-70b. 92 | 93 | * Reference Model: Defined using --ref_model. Example: --ref_model huggyllama/llama-7b. 94 | 95 | * Data Length: Define the length for the WikiMIA benchmark with --length. Available options: 32, 54, 128, 256. 96 | 97 | 📌 Note: ***For optimal results, use fixed-length inputs with our Min-K% Prob method*** (When you evalaute Min-K% Prob method on your own dataset, make sure the input length of each example is the same.) 98 | 99 | 📊 Baselines: Our script comes with the following baselines: PPL, Calibration Method, PPL/zlib_compression, PPL/lowercase_ppl 100 | 101 | -------------------------------------------------------------------------------- /mink_prob.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/swj0419/detect-pretrain-code/a5048f93c4fa6491aab0638bb3e6fd5a662d7698/mink_prob.png -------------------------------------------------------------------------------- /process_data.py: -------------------------------------------------------------------------------- 1 | from datasets import load_dataset 2 | from datasets import Dataset 3 | from ipdb import set_trace as bp 4 | # from options import Options 5 | from src.eval import * 6 | 7 | 8 | # data: different length: 9 | def process_each_dict_length_data(data, length=32): 10 | new_data = [] 11 | for ex in data: 12 | ex_copy = ex.copy() 13 | if len(ex_copy["input"].split()) < length: 14 | continue 15 | else: 16 | ex_copy["input"] = " ".join(ex_copy["input"].split()[:length]) 17 | new_data.append(ex_copy) 18 | return new_data 19 | 20 | def change_type(data): 21 | new_data = {"input": [], "label": []} 22 | for ex in data: 23 | ex["label"] = int(ex["label"]) 24 | new_data["input"].append(ex["input"]) 25 | new_data["label"].append(ex["label"]) 26 | return new_data 27 | 28 | if __name__ == "__main__": 29 | dataset = load_jsonl("/fsx-onellm/swj0419/attack/detect-pretrain-code/data/wikimia.jsonl") 30 | # bp() 31 | data = convert_huggingface_data_to_list_dic(dataset) 32 | for length in [128, 256]: 33 | new_data = process_each_dict_length_data(data, length=length) 34 | print(f"length {length} data size: {len(new_data)}") 35 | dump_jsonl(new_data, f"data/WikiMIA_length{length}.jsonl") 36 | huggingface_dataset = Dataset.from_dict(change_type(new_data)) 37 | huggingface_dataset.push_to_hub("swj0419/WikiMIA", f"WikiMIA_length{length}") 38 | 39 | 40 | -------------------------------------------------------------------------------- /src/__pycache__/eval.cpython-39.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/swj0419/detect-pretrain-code/a5048f93c4fa6491aab0638bb3e6fd5a662d7698/src/__pycache__/eval.cpython-39.pyc -------------------------------------------------------------------------------- /src/__pycache__/options.cpython-39.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/swj0419/detect-pretrain-code/a5048f93c4fa6491aab0638bb3e6fd5a662d7698/src/__pycache__/options.cpython-39.pyc -------------------------------------------------------------------------------- /src/eval.py: -------------------------------------------------------------------------------- 1 | import logging 2 | logging.basicConfig(level='ERROR') 3 | import numpy as np 4 | from tqdm import tqdm 5 | import json 6 | from collections import defaultdict 7 | import matplotlib.pyplot as plt 8 | from sklearn.metrics import auc, roc_curve 9 | import matplotlib 10 | import random 11 | 12 | 13 | matplotlib.rcParams['pdf.fonttype'] = 42 14 | matplotlib.rcParams['ps.fonttype'] = 42 15 | 16 | 17 | matplotlib.rcParams['pdf.fonttype'] = 42 18 | matplotlib.rcParams['ps.fonttype'] = 42 19 | 20 | # plot data 21 | def sweep(score, x): 22 | """ 23 | Compute a ROC curve and then return the FPR, TPR, AUC, and ACC. 24 | """ 25 | fpr, tpr, _ = roc_curve(x, -score) 26 | acc = np.max(1-(fpr+(1-tpr))/2) 27 | return fpr, tpr, auc(fpr, tpr), acc 28 | 29 | 30 | def do_plot(prediction, answers, sweep_fn=sweep, metric='auc', legend="", output_dir=None): 31 | """ 32 | Generate the ROC curves by using ntest models as test models and the rest to train. 33 | """ 34 | fpr, tpr, auc, acc = sweep_fn(np.array(prediction), np.array(answers, dtype=bool)) 35 | 36 | low = tpr[np.where(fpr<.05)[0][-1]] 37 | # bp() 38 | print('Attack %s AUC %.4f, Accuracy %.4f, TPR@5%%FPR of %.4f\n'%(legend, auc,acc, low)) 39 | 40 | metric_text = '' 41 | if metric == 'auc': 42 | metric_text = 'auc=%.3f'%auc 43 | elif metric == 'acc': 44 | metric_text = 'acc=%.3f'%acc 45 | 46 | plt.plot(fpr, tpr, label=legend+metric_text) 47 | return legend, auc,acc, low 48 | 49 | 50 | def fig_fpr_tpr(all_output, output_dir): 51 | print("output_dir", output_dir) 52 | answers = [] 53 | metric2predictions = defaultdict(list) 54 | for ex in all_output: 55 | answers.append(ex["label"]) 56 | for metric in ex["pred"].keys(): 57 | if ("raw" in metric) and ("clf" not in metric): 58 | continue 59 | metric2predictions[metric].append(ex["pred"][metric]) 60 | 61 | plt.figure(figsize=(4,3)) 62 | with open(f"{output_dir}/auc.txt", "w") as f: 63 | for metric, predictions in metric2predictions.items(): 64 | legend, auc, acc, low = do_plot(predictions, answers, legend=metric, metric='auc', output_dir=output_dir) 65 | f.write('%s AUC %.4f, Accuracy %.4f, TPR@0.1%%FPR of %.4f\n'%(legend, auc, acc, low)) 66 | 67 | plt.semilogx() 68 | plt.semilogy() 69 | plt.xlim(1e-5,1) 70 | plt.ylim(1e-5,1) 71 | plt.xlabel("False Positive Rate") 72 | plt.ylabel("True Positive Rate") 73 | plt.plot([0, 1], [0, 1], ls='--', color='gray') 74 | plt.subplots_adjust(bottom=.18, left=.18, top=.96, right=.96) 75 | plt.legend(fontsize=8) 76 | plt.savefig(f"{output_dir}/auc.png") 77 | 78 | 79 | def load_jsonl(input_path): 80 | with open(input_path, 'r') as f: 81 | data = [json.loads(line) for line in tqdm(f)] 82 | random.seed(0) 83 | random.shuffle(data) 84 | return data 85 | 86 | def dump_jsonl(data, path): 87 | with open(path, 'w') as f: 88 | for line in tqdm(data): 89 | f.write(json.dumps(line) + "\n") 90 | 91 | def read_jsonl(path): 92 | with open(path, 'r') as f: 93 | return [json.loads(line) for line in tqdm(f)] 94 | 95 | def convert_huggingface_data_to_list_dic(dataset): 96 | all_data = [] 97 | for i in range(len(dataset)): 98 | ex = dataset[i] 99 | all_data.append(ex) 100 | return all_data -------------------------------------------------------------------------------- /src/options.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import os 3 | from pathlib import Path 4 | import logging 5 | 6 | logger = logging.getLogger(__name__) 7 | 8 | class Options(): 9 | def __init__(self): 10 | self.parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter) 11 | self.initialize_parser() 12 | 13 | def initialize_parser(self): 14 | self.parser.add_argument('--target_model', type=str, default="text-davinci-003", help="the model to attack: huggyllama/llama-65b, text-davinci-003") 15 | self.parser.add_argument('--ref_model', type=str, default="huggyllama/llama-7b") 16 | self.parser.add_argument('--output_dir', type=str, default="out") 17 | self.parser.add_argument('--data', type=str, default="swj0419/WikiMIA", help="the dataset to evaluate: default is WikiMIA") 18 | self.parser.add_argument('--length', type=int, default=64, help="the length of the input text to evaluate. Choose from 32, 64, 128, 256") 19 | self.parser.add_argument('--key_name', type=str, default="input", help="the key name corresponding to the input text. Selecting from: input, parapgrase") 20 | 21 | 22 | 23 | 24 | -------------------------------------------------------------------------------- /src/out/huggyllama/llama-13b_huggyllama/llama-7b/input/auc.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/swj0419/detect-pretrain-code/a5048f93c4fa6491aab0638bb3e6fd5a662d7698/src/out/huggyllama/llama-13b_huggyllama/llama-7b/input/auc.png -------------------------------------------------------------------------------- /src/out/huggyllama/llama-13b_huggyllama/llama-7b/input/auc.txt: -------------------------------------------------------------------------------- 1 | ppl AUC 0.5865, Accuracy 0.6090, TPR@0.1%FPR of 0.2308 2 | ppl/Ref_ppl (calibrate PPL to the reference model) AUC 0.7019, Accuracy 0.7163, TPR@0.1%FPR of 0.4231 3 | ppl/lowercase_ppl AUC 0.5753, Accuracy 0.5994, TPR@0.1%FPR of 0.1538 4 | ppl/zlib AUC 0.7179, Accuracy 0.7179, TPR@0.1%FPR of 0.3077 5 | Min_20.0% Prob AUC 0.6186, Accuracy 0.6474, TPR@0.1%FPR of 0.1538 6 | Min_30.0% Prob AUC 0.6106, Accuracy 0.6250, TPR@0.1%FPR of 0.1538 7 | Min_40.0% Prob AUC 0.5913, Accuracy 0.5929, TPR@0.1%FPR of 0.1923 8 | Min_50.0% Prob AUC 0.5833, Accuracy 0.6106, TPR@0.1%FPR of 0.2308 9 | Min_60.0% Prob AUC 0.5849, Accuracy 0.6106, TPR@0.1%FPR of 0.2308 10 | -------------------------------------------------------------------------------- /src/out/huggyllama/llama-13b_huggyllama/llama-7b/input/auc_google.txt: -------------------------------------------------------------------------------- 1 | ppl 0.663 2 | ppl/Ref_ppl (calibrate PPL to the reference model) 0.626 3 | ppl/lowercase_ppl 0.58 4 | ppl/zlib 0.619 5 | mean 0.663 6 | Min_0.2% Prob 0.677 7 | Min_0.3% Prob 0.678 8 | Min_0.4% Prob 0.675 9 | Min_0.5% Prob 0.67 10 | Min_0.6% Prob 0.666 11 | -------------------------------------------------------------------------------- /src/out/huggyllama/llama-13b_huggyllama/llama-7b/input/low_google.txt: -------------------------------------------------------------------------------- 1 | ppl 0.086 2 | ppl/Ref_ppl (calibrate PPL to the reference model) 0.074 3 | ppl/lowercase_ppl 0.069 4 | ppl/zlib 0.137 5 | mean 0.086 6 | Min_0.2% Prob 0.127 7 | Min_0.3% Prob 0.104 8 | Min_0.4% Prob 0.096 9 | Min_0.5% Prob 0.089 10 | Min_0.6% Prob 0.089 11 | -------------------------------------------------------------------------------- /src/out/text-davinci-003_huggyllama/llama-7b/input/auc.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/swj0419/detect-pretrain-code/a5048f93c4fa6491aab0638bb3e6fd5a662d7698/src/out/text-davinci-003_huggyllama/llama-7b/input/auc.png -------------------------------------------------------------------------------- /src/out/text-davinci-003_huggyllama/llama-7b/input/auc.txt: -------------------------------------------------------------------------------- 1 | ppl AUC 0.7420, Accuracy 0.7452, TPR@0.1%FPR of 0.3846 2 | ppl/Ref_ppl (calibrate PPL to the reference model) AUC 0.2949, Accuracy 0.5176, TPR@0.1%FPR of 0.0769 3 | ppl/lowercase_ppl AUC 0.6747, Accuracy 0.6747, TPR@0.1%FPR of 0.3077 4 | ppl/zlib AUC 0.7564, Accuracy 0.7356, TPR@0.1%FPR of 0.4231 5 | Min_20.0% Prob AUC 0.7724, Accuracy 0.8237, TPR@0.1%FPR of 0.4231 6 | Min_30.0% Prob AUC 0.7660, Accuracy 0.8237, TPR@0.1%FPR of 0.3846 7 | Min_40.0% Prob AUC 0.7660, Accuracy 0.8029, TPR@0.1%FPR of 0.4231 8 | Min_50.0% Prob AUC 0.7468, Accuracy 0.7644, TPR@0.1%FPR of 0.3846 9 | Min_60.0% Prob AUC 0.7420, Accuracy 0.7452, TPR@0.1%FPR of 0.3846 10 | -------------------------------------------------------------------------------- /src/run.py: -------------------------------------------------------------------------------- 1 | import logging 2 | logging.basicConfig(level='ERROR') 3 | import numpy as np 4 | from pathlib import Path 5 | import openai 6 | import torch 7 | import zlib 8 | from transformers import AutoTokenizer, AutoModelForCausalLM 9 | from tqdm import tqdm 10 | import numpy as np 11 | from datasets import load_dataset 12 | from options import Options 13 | from eval import * 14 | 15 | 16 | def load_model(name1, name2): 17 | if "davinci" in name1: 18 | model1 = None 19 | tokenizer1 = None 20 | else: 21 | model1 = AutoModelForCausalLM.from_pretrained(name1, return_dict=True, device_map='auto') 22 | model1.eval() 23 | tokenizer1 = AutoTokenizer.from_pretrained(name1) 24 | 25 | if "davinci" in name2: 26 | model2 = None 27 | tokenizer2 = None 28 | else: 29 | model2 = AutoModelForCausalLM.from_pretrained(name2, return_dict=True, device_map='auto') 30 | model2.eval() 31 | tokenizer2 = AutoTokenizer.from_pretrained(name2) 32 | return model1, model2, tokenizer1, tokenizer2 33 | 34 | def calculatePerplexity_gpt3(prompt, modelname): 35 | prompt = prompt.replace('\x00','') 36 | responses = None 37 | # Put your API key here 38 | openai.api_key = "YOUR_API_KEY" # YOUR_API_KEY 39 | while responses is None: 40 | try: 41 | responses = openai.Completion.create( 42 | engine=modelname, 43 | prompt=prompt, 44 | max_tokens=0, 45 | temperature=1.0, 46 | logprobs=5, 47 | echo=True) 48 | except openai.error.InvalidRequestError: 49 | print("too long for openai API") 50 | data = responses["choices"][0]["logprobs"] 51 | all_prob = [d for d in data["token_logprobs"] if d is not None] 52 | p1 = np.exp(-np.mean(all_prob)) 53 | return p1, all_prob, np.mean(all_prob) 54 | 55 | 56 | def calculatePerplexity(sentence, model, tokenizer, gpu): 57 | """ 58 | exp(loss) 59 | """ 60 | input_ids = torch.tensor(tokenizer.encode(sentence)).unsqueeze(0) 61 | input_ids = input_ids.to(gpu) 62 | with torch.no_grad(): 63 | outputs = model(input_ids, labels=input_ids) 64 | loss, logits = outputs[:2] 65 | 66 | ''' 67 | extract logits: 68 | ''' 69 | # Apply softmax to the logits to get probabilities 70 | probabilities = torch.nn.functional.log_softmax(logits, dim=-1) 71 | # probabilities = torch.nn.functional.softmax(logits, dim=-1) 72 | all_prob = [] 73 | input_ids_processed = input_ids[0][1:] 74 | for i, token_id in enumerate(input_ids_processed): 75 | probability = probabilities[0, i, token_id].item() 76 | all_prob.append(probability) 77 | return torch.exp(loss).item(), all_prob, loss.item() 78 | 79 | 80 | def inference(model1, model2, tokenizer1, tokenizer2, text, ex, modelname1, modelname2): 81 | pred = {} 82 | 83 | if "davinci" in modelname1: 84 | p1, all_prob, p1_likelihood = calculatePerplexity_gpt3(text, modelname1) 85 | p_lower, _, p_lower_likelihood = calculatePerplexity_gpt3(text.lower(), modelname1) 86 | else: 87 | p1, all_prob, p1_likelihood = calculatePerplexity(text, model1, tokenizer1, gpu=model1.device) 88 | p_lower, _, p_lower_likelihood = calculatePerplexity(text.lower(), model1, tokenizer1, gpu=model1.device) 89 | 90 | if "davinci" in modelname2: 91 | p_ref, all_prob_ref, p_ref_likelihood = calculatePerplexity_gpt3(text, modelname2) 92 | else: 93 | p_ref, all_prob_ref, p_ref_likelihood = calculatePerplexity(text, model2, tokenizer2, gpu=model2.device) 94 | 95 | # ppl 96 | pred["ppl"] = p1 97 | # Ratio of log ppl of large and small models 98 | pred["ppl/Ref_ppl (calibrate PPL to the reference model)"] = p1_likelihood-p_ref_likelihood 99 | 100 | 101 | # Ratio of log ppl of lower-case and normal-case 102 | pred["ppl/lowercase_ppl"] = -(np.log(p_lower) / np.log(p1)).item() 103 | # Ratio of log ppl of large and zlib 104 | zlib_entropy = len(zlib.compress(bytes(text, 'utf-8'))) 105 | pred["ppl/zlib"] = np.log(p1)/zlib_entropy 106 | # min-k prob 107 | for ratio in [0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6]: 108 | k_length = int(len(all_prob)*ratio) 109 | topk_prob = np.sort(all_prob)[:k_length] 110 | pred[f"Min_{ratio*100}% Prob"] = -np.mean(topk_prob).item() 111 | 112 | ex["pred"] = pred 113 | return ex 114 | 115 | def evaluate_data(test_data, model1, model2, tokenizer1, tokenizer2, col_name, modelname1, modelname2): 116 | print(f"all data size: {len(test_data)}") 117 | all_output = [] 118 | test_data = test_data 119 | for ex in tqdm(test_data): 120 | text = ex[col_name] 121 | new_ex = inference(model1, model2, tokenizer1, tokenizer2, text, ex, modelname1, modelname2) 122 | all_output.append(new_ex) 123 | return all_output 124 | 125 | 126 | if __name__ == '__main__': 127 | args = Options() 128 | args = args.parser.parse_args() 129 | args.output_dir = f"{args.output_dir}/{args.target_model}_{args.ref_model}/{args.key_name}" 130 | Path(args.output_dir).mkdir(parents=True, exist_ok=True) 131 | 132 | # load model and data 133 | model1, model2, tokenizer1, tokenizer2 = load_model(args.target_model, args.ref_model) 134 | if "jsonl" in args.data: 135 | data = load_jsonl(f"{args.data}") 136 | else: # load data from huggingface 137 | dataset = load_dataset(args.data, split=f"WikiMIA_length{args.length}") 138 | data = convert_huggingface_data_to_list_dic(dataset) 139 | 140 | all_output = evaluate_data(data, model1, model2, tokenizer1, tokenizer2, args.key_name, args.target_model, args.ref_model) 141 | fig_fpr_tpr(all_output, args.output_dir) 142 | 143 | --------------------------------------------------------------------------------