Legal AI has a wide range of applications such as predicting whether a prosecution will be punished, or whether the punishment will be a prison sentence or a fine. However, current advances in natural language processing have resulted in an ever-increasing number of language models. The cost of finetuning the pre-trained language model and storing these fine-tuned language models becomes more and more expensive. To address this issue, we adopted the concept of Parameter Eficient Fine-Tuning (PEFT) and applied it to the field of Legal AI. By leveraging PEFT techniques, particularly through the implementation of the Low-Rank Adaptation (LoRA) architecture, we have achieved promising results in fine-tuning pre-trained language models. This approach enables us to achieve comparable, if not superior, performance while significantly reducing the time required for model adjustments. It demonstrates the potential of PEFT techniques in adapting language models to diferent legal frameworks, enhancing the accuracy and relevance of legal knowledge services, and making Legal AI more accessible to individuals without legal backgrounds.
Legal AI refers to the utilization of artificial intelligence (AI) technology in the legal sector. It
is an expanding field that harnesses sophisticated algorithms and machine learning techniques
to assist in the organization, analysis, and interpretation of extensive legal documentation.
Applications of legal AI encompass various areas, including case management [
Legal cases typically fall into two main categories: civil law and criminal law. Since
gathering facts and evidence for civil cases can be challenging [
In recent years, significant progress has also been made on many legal tasks based on
pretrained models, including accusation prediction [9], prison term classification [
In this study, we propose the use of PEFT to fine-tune pre-trained language models. Specifically, we adopt Low-Rank Adaptation of Large Language Models (LoRA)[10], as an implementation of PEFT, which ofers advantages in reducing computational resources and fine-tuning time while maintaining or surpassing model performance. This makes it particularly valuable for refining large models with billions of parameters.
The rest of the paper is organized as follows: Section 2 introduces related work on legal AI and LJP and Parameter-Eficient Fine-Tuning (PEFT). The problem definition and dataset construction is detailed in Section 3. Section 4 explains PEFT. We report the experimental results in Section 5. Finally, Section 6 concludes the paper and suggests for future research direction.
Legal artificial intelligence (LegalAI) has drawn increasing attention from NLP researchers because of the vast amount of legal documents. Zhong et al. [11] surveyed the researches on legal artificial intelligence (LegalAI) and categorized its applications into three types: legal judgment prediction (LJP), similar case matching, and legal question answering.
Among them, legal judgment prediction has been widely studied for decades, and there are
also several related LJP datasets, such as CAIL [
LegalAI’s research methods can be divided into symbol-based methods and embedding-based
methods [11]. In the past, researchers have used traditional machine learning methods for
feature extraction, attempting to extract or create specific features from the description of criminal
facts using additional labeling to help describe the crime. For example, Hu et al. [
To address the above scaling issues, researchers have attempted to incorporate legal
knowledge into neural networks via automatic learning. For example, Luo et al. [17] adopted a
twostep approach to filter out irrelevant law articles with and retain the top k articles to scale up
to a large number of law articles. They built a binary classifier for each article focusing on
its relevance to the input case. The advantage of such an approach is that we can add new
articles with the existing classifiers untouched. Similarly, Bao et al. [ 9] proposed an attention
neural network, LegalAtt, which uses relevant articles to improve the performance and
interpretability of charge prediction task. Gan et al. [18] injected the legal knowledge in the form
as a set of first-order logic rules and integrate these rules into a co-attention network-based
model, which makes the prediction more interpretable for civil loan cases. Kang et al. [16]
constructed auxiliary fact representations from the definitions of behavioral reasons to enhance
fact descriptions. Lyu et al. [
Multi-task learning framework is a machine learning method that can train multiple related tasks simultaneously, thus improving the performance of each task. It can use a shared layer to extract common features for all tasks, and then use diferent specialized layers to handle the details of each task, or use diferent layers to extract features for each task and then use some methods to limit the diferences between the parameters of these layers. Zhong et al. [7] proposed the TopJudge model, which uses a topological graph to enhance performance by exploiting the relationships between legal judgments, predicting articles, charges, and terms. Yang et al. [19] proposed a multi-layer forward prediction and backward validation framework to efectively utilize the dependency relationships between multiple sub-tasks.
It has been shown that it is feasible to update or add a very small number of parameters as opposed to updating all of the parameters of the pre-trained model as is the case with ordinary ifne-tuning. The addition of adapters, which are tiny trainable feed-forward networks inserted between the layers of the fixed pre-trained model, was suggested Houlsby et al. [ 20] (See Figure 1). Since then, a wide range of advanced PEFT techniques have been put forth, e.g. low rank adaptation by Hu et al. [10], and prefix-tuning by [ 21]. In a way, Houlsby et al. (2019) places two adapters sequentially within one layer of the transformer, that is to say typical adapters are sequential computation. On the other hand, prefix-tuning and LoRA can be thought of as a “parallel”computation to the PLM layer. An unified view toward Parameter-eficient transfer learning was proposed by He et al. [22]. 3. Problem Formulation and Dataset Construction Four steps make up a criminal proceeding: investigation, prosecution, trial, and execution. Among these steps, the public is most interested in the investigation and trial steps. The investigation procedure refers to the process in which law enforcement agents look into potential criminal events and gather evidence under the direction of the prosecutor. The prosecution will file charges and begin the trial process if they feel that the defendant has a strong suspicion of committing a crime. An impartial, unbiased judge oversees the trial process and determines whether the defendant actually committed a crime based on the evidence given by the prosecutor. Today, judgment documents are used as the data source in the majority of publicly accessible datasets for LJP research. However the language employed in judgment documents is frequently more eloquent, and the substance primarily concentrates on the facts and procedures, leading to greater document lengths and more dificult comprehension for legal specialists. On the other hand, prosecutors employ language that is shorter and more akin to that of the general public when describing the portion of the criminal facts in the indictment that are based on their involvement in the investigation. Hence, rather than using judgment documents for the scope of the data collection, we employ indictments.
We collected indictments from the public document inquiry system of the Ministry of Justice of Taiwan from June 15, 2018 to June 30, 2021. The defendant, charges, criminal facts, and legal provisions were extracted from the indictments using regular expressions, and the material was then organized into a JSON format. There were 533 articles under 41 laws and 183 charges from 355,295 cases in the original dataset.
How many articles and charges to include in the prediction model is a recurring issue while creating the LJP dataset. We screened out instances where the number of charges or articles was insuficient in order to make the experiment fair and prevent classification-related insuficient training or testing data, which may have an impact on the experimental outcomes (e.g., less than 30 cases). Furthermore, the first 100 articles of Taiwan’s criminal code contain definitions of terms like attempted ofenses and criminal responsibility, but we did not include these articles in our dataset because they do not specify the real penalties. Excluding the above cases, the total number of articles decreased significantly to 165, and the number of charges decreased from 183 to 94. A total of 12,541 cases were removed, accounting for 3.5% of the total dataset. It is worth noting that a case may violate more than one charge, but often only the primary Facts Laws Articles ⋯Knowing that handing over account passbooks, financial cards, and passwords to oth⋯知悉將帳戶存簿、金融卡及密碼交 ers may serve as tools for criminals to com付他人使用，恐為不法者充作詐騙被 mit fraud by transferring funds, and also not 害人匯入款項之犯罪工具，亦不違背 deviating from their intention of money laun其本意之洗錢及幫助詐欺取財之犯意， dering and aiding in fraudulent schemes, pro將其之存摺及提款卡等資料，並提供 viding the passbooks, withdrawal cards, and 提款卡密碼，以寄送包裹之方式，租借 supplying the PIN codes through parcel de寄予詐欺集團成員，容任該人及其所 livery to members of a fraudulent group en屬之詐騙集團持以犯罪使用。⋯ ables that person and their afiliated fraudulent organization to utilize them for criminal purposes. ⋯
Money Laundering Control Act、Criminal 洗錢防制法、刑法 Code
It was committed by the defendant, and is guilty of the crime of money laundering un⋯是核被告所為，係犯洗錢防制法第 2 der Article 2, paragraph 2, and Article 14, para條第 2 款、第 14 條第 1 項之洗錢罪嫌 graph 1, and the crime of assisting in fraud及刑法第 30 條第 1 項前段、第 339 條 ulent acquisition of money under Article 30, 第 1 項之幫助詐欺取財罪嫌⋯ paragraph 1, and Article 339, paragraph 1 of the Criminal Code.
詐欺 Fraud charge are listed in the indictment. Thus, it is more dificult to estimate charge than articles (even though the number of articles in our dataset is greater than the number of charges). The distribution of instances in this dataset is unequal, as one might anticipate. The top 10 counts make about 85% of all cases, according to the number of charges in the indictment. In contrast, just 0.14% of the instances are covered by the lowest 10 charges. We divided the cases into categories based on the charges in the indictment in order to fairly split the data, using 80% of the instances in each category as training data, 10% as validation data, and the final 10% as testing data. Lastly, we created a dataset called TWLJP1 (TaiWan Legal Judgment Prediction Datasets) by combining the data from all categories to create training, validation, and testing datasets. blue and green, respectively. In this instance, a suspect gave the fraudsters access to his bank account, and the group tricked the victim into wiring money to the account before withdrawing it. The Anti-Money Laundering Act and the Criminal Law were both allegedly broken by the defendant, however the indictment only listed fraud as a crime.
Dataset TWLJP # cases 342,754 # laws # articles # charge 33 165 94 avg length
Let = ( sequence of 1, 2, ⋯ , ) denotes a dataset with cases where each case is described by a words
= ( 1, 2, ⋯ , ), and is associated with three labels in , in and in , where and denote the size of the one-hot vector of law and charge, while is of the three vectors and , ,
in {0, 1}. a multi-hot vector of articles with dimension . Each case is also associated with a vector of laws, = ( 1 , 2, ⋯ , ), a vector of articles, = ( 1
, 2, ⋯ , ), and a vector of charges, = ( 1
, 2, ⋯ , ), where , , represent the size
Current models like Lawformer and TopJudge, as well as other state-of-the-art Legal Judgment Prediction (LJP) models, showcase the potential of neural network models in terms of accuracy and eficiency in predicting legal judgments. However, it is important to acknowledge that these models have certain limitations when applied to legal systems of diferent countries. For example, Lawformer is a pre-trained language model that utilizes legal documents from Mainland China as training data. It has shown impressive performance on the CAIL dataset. However, when applied to the TWLJP dataset, its performance is not as good as Chinese BERT. The reason could be attributed to variations in legal terminology, penalties, and writing styles of legal documents across diferent countries.
As mentioned before, Parameter Eficient Fine-Tuning (PEFT) is an alternative approach that allows a model to learn a new task with minimal updates. In PEFT, a pre-trained model is finetuned by selectively updating or adding a small number of parameters. Recent advancements in PEFT techniques have demonstrated the ability to achieve performance comparable to finetuning the entire model while only modifying a fraction (e.g., 0.01%) of its parameters [23].
In this paper, we adopt LoRA[10] to reduce the number of trainable parameters by learning pairs of rank-decomposition matrices while keeping the original weights frozen. The idea behind LoRA is that when adapting a pre-trained language model to a specific task or dataset, only a few features need to be emphasized or re-learnt. This means that the update matrix (ΔW) can be a low-rank matrix. As shown in Figure 2, The update of a pre-trained weight matrix ∈ × is constrained by using a low-rank decomposition 0 + ∆ = 0 + , where ∈ × , ∈ × , and the rank r is a hyper parameter less than or equal to the minimum of d and k. During training, 0 remains unchanged and does not receive any gradient updates, while A and B contain trainable parameters.
Since our LJP tasks include the prediction of legal law, article, and charges, we add three fully connected layers to the CLS output as depicted in Figure 2. By adopting this approach with PLM frozen, we can significantly minimize the computational resources and time required for fine-tuning while ensuring the model’s performance is preserved.
The key advantage of LoRA lies in its remarkable ability to substantially reduce the computational resources and time necessary for fine-tuning, while maintaining the model’s performance. This method proves particularly valuable when tackling extensive fine-tuning tasks, such as the refinement of highly capable large models that consist of billions of parameters.
In order to evaluate the performance of the TWLJP dataset that we have collected across different pre-trained language models, we conducted training and evaluation using the following settings: Multi-task BERT As dipicted in Figure2, we use multi-task learning to model the prediction of Law, Charge and Article by given criminal fact descriptions in the indictment as input. We utilize the Huggingface[24] Chinese pre-training language model bert-base-chinese. The optimizer we use for Multi-task BERT is BERT Adam with a learning rate of 1e-5, maximum length of 512 and hidden size of 768 for the parameters of pre-trained language model.
Multi-task Lawformer Lawformer[12] is a pre-trained language model based on the CAILlong dataset and capable of processing articles up to 4096 characters in length. However, since Lawformer uses the CAIL-long dataset in simplified Chinese, and our data is in traditional Chinese, we first used the OpenCC package to convert the crime facts to simplified Chinese before training. The optimizer we use for Multi-task Lawformer is AdamW with a learning rate of 1e-5, maximum length of 512 and hidden size of 768 for the parameters of pre-trained language model.
LoRA We utilized the LoRA implementation from Hugging Face’s PEFT package[25] and bertbased-chinese model to generate embeddings. In the LoRA setting, the value of r is set to 8. The optimizer we use for LoRA is AdamW with a learning rate of 3e-4, maximum length of 512 and hidden size of 768 for the parameters of pre-trained language model.
Evaluation Metric We adopt micro precision (MiP), recall (MiR), and F1 score (MiF), as well as macro precision (MaP), recall (MaR), and F1 score (MaF), as the evaluation metrics. Macroprecision/recall/F1 is computed by averaging each class, which is a commonly used metric in multi-label classification tasks.
To evaluate the performance of the TWLJP dataset across diferent models, we conducted experiments using the models introduced in the previous section. The performance of TWLJP on each model is shown in Tables 3, 4, and 5. In each experiment, we selected the epoch with the best performance on the validation dataset and tested on the testing dataset. The performance shown in the tables is the average performance of the model over five experiments, with a calculation of 2 times the standard deviation.
We conducted the experiments using the GeForce RTX 4070 Ti graphics card, and the training time for each model for one epoch, as well as the parameter information of the models, are presented in Table 6.
Based on the experimental results, it is evident that the performance of models implemented using the Lawformer pre-trained language model did not meet our expectations. Upon analysis, we determined that the reason behind this discrepancy lies in the fact that Lawformer was trained on legal documents from mainland China. Despite our eforts to convert the input criminal facts from Traditional Chinese to Simplified Chinese, there are significant diferences between the legal systems and terminologies used in mainland China and Taiwan. This mismatch in legal terminology and usage negatively impacted the performance of Lawformer on the TWLJP dataset.
Under the training architecture of LoRA, comparable performance to Multi-task BERT is achieved in terms of case cause, legal provisions, and legal sources, and even superior performance compared to Multi-task BERT. The training time for one epoch is 1 hour and 58 minutes, which is approximately half the time required by Multi-task BERT, which is 3 hours and 49 minutes. Regarding the parameter count, Multi-task BERT has a total of 102,716,744 parameters, all of which need adjustment. In the LoRA architecture, the total number of parameters is 103,011,656, but only 744,008 parameters need to be trained, which is approximately 0.72% of the trainable parameters in Multi-task BERT.
Sub-task
Model/Metric
Multi-task BERT Multi-task Lawformer
LoRA(r=8)
MiP 84.1 79.3 84.6
MiR 85.7 79.8 86.9
Article MiF MaP 84.9 79.0 79.6 70.9 85.8 79.9
MaR 71.6 59.6 71.6
MaF 73.4 62.7 73.9
To ensure fairness in our experiments, we also utilized the publicly available CAIL dataset. We conducted multi-task training on the dataset, focusing on the charges and articles. The perfor
Sub-task
Model/Metric
Multi-task BERT Multi-task Lawformer
LoRA(r=8) mance of each model is shown in Tables 7 and 8. Since the main objective of our experiment was to compare the performance, time, and parameters of large language models, we did not compare them to other related models. For each experiment, we selected the epoch with the best performance on the validation dataset and tested it on the test dataset. We conducted the experiments using the GeForce RTX 4070 Ti graphics card, and the training time for each model per epoch and the parameter information are provided in Table 9.
From the experimental results, it can be observed that the performance of the Lawformer pretrained language model did not meet expectations. Upon analyzing the reasons for this, although Lawformer was trained on legal documents from mainland China, it is based on the Longformer architecture, which allows for input lengths of up to 4096 tokens. However, we used a maximum length of 512 tokens, and modifying this maximum length would lead to insuficient memory on the graphics card. As a result, the weights of some models were not updated, leading to poor training performance.
On the other hand, under the training framework of LoRA, comparable performance to Multitask BERT was achieved for charges and articles. The training time for one epoch was 1 hour and 14 minutes, compared to 2 hours and 19 minutes for Multi-task BERT, requiring approximately half the time. In terms of parameter quantity, LoRA only required approximately 0.86% of the training parameters compared to Multi-task BERT.
Legal AI plays a crucial role in providing legal knowledge services to individuals with legal backgrounds, as well as assisting non-legal professionals. However, due to the diverse range of sub-tasks in Legal AI and the increasing size of pre-trained language models, training and storing a separate language model for each sub-task can be costly and resource-intensive. To address this challenge, we have embraced the concept of Parameter Eficient Fine-Tuning (PEFT) and applied it to the field of Legal AI.
By leveraging the PEFT approach, specifically through the implementation of the LoRA architecture, we have observed promising results in fine-tuning pre-trained language models. This approach allows us to achieve comparable, if not superior, performance while significantly reducing the time required for model adjustments. In our experiments, we found that using the LoRA framework required only about half the time compared to fine-tuning the entire model, without sacrificing performance. This innovative methodology opens up new possibilities for adapting language models to diferent legal contexts eficiently.
The success of our approach highlights the potential of PEFT techniques in the Legal AI domain. By eficiently adjusting and fine-tuning language models, we can tailor them to specific legal frameworks, taking into account the variations in legal definitions, documents, and terminologies across diferent countries. This advancement not only enhances the accuracy and relevance of legal knowledge services but also extends the accessibility of Legal AI to individuals without a legal background. [7] H. Zhong, Z. Guo, C. Tu, C. Xiao, Z. Liu, M. Sun, Legal judgment prediction via topological learning, in: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Brussels, Belgium, 2018, pp. 3540–3549. URL: https://aclanthology.org/D18-1390. doi:10.18653/v1/D18- 1390. [8] N. Xu, P. Wang, L. Chen, L. Pan, X. Wang, J. Zhao, Distinguish confusing law articles for legal judgment prediction, in: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics, Online, 2020, pp. 3086–3095. URL: https://aclanthology.org/2020.acl-main.280. doi:10.18653/v1/2020.acl- main.280. [9] Q. Bao, H. Zan, P. Gong, J. Chen, Y. Xiao, Charge prediction with legal attention, in: Natural Language Processing and Chinese Computing: 8th CCF International Conference, NLPCC 2019, Dunhuang, China, October 9–14, 2019, Proceedings, Part I, Springer-Verlag, Berlin, Heidelberg, 2019, p. 447–458. URL: https://doi.org/10.1007/978-3-030-32233-5_35. doi:10.1007/978- 3- 030- 32233- 5_35. [10] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, Lora: Low-rank adaptation of large language models, arXiv preprint arXiv:2106.09685 (2021). [11] H. Zhong, C. Xiao, C. Tu, T. Zhang, Z. Liu, M. Sun, How does NLP benefit legal system: A summary of legal artificial intelligence, in: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics, Online, 2020, pp. 5218–5230. URL: https://aclanthology.org/2020.acl-main.466. doi:10.18653/v1/2020.acl- main.466. [12] C. Xiao, X. Hu, Z. Liu, C. Tu, M. Sun, Lawformer: A pre-trained language model for chinese legal long documents, AI Open 2 (2021) 79–84. URL: https://www.sciencedirect. com/science/article/pii/S2666651021000176. doi:10.1016/j.aiopen.2021.06.003. [13] N. Aletras, D. Tsarapatsanis, D. Preoţiuc-Pietro, V. Lampos, Predicting judicial decisions of the european court of human rights: A natural language processing perspective, PeerJ Computer Science 2 (2016) e93. [14] I. Chalkidis, I. Androutsopoulos, N. Aletras, Neural legal judgment prediction in English, in: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics, Florence, Italy, 2019, pp. 4317–4323.
URL: https://aclanthology.org/P19-1424. doi:10.18653/v1/P19- 1424. [15] R. A. Shaikh, T. P. Sahu, V. Anand, Predicting outcomes of legal cases based on legal factors using classifiers, Procedia Computer Science 167 (2020) 2393–2402. URL: https:// www.sciencedirect.com/science/article/pii/S1877050920307584. doi:10.1016/j.procs. 2020.03.292, international Conference on Computational Intelligence and Data Science. [16] L. Kang, J. Liu, L. Liu, D. Ye, Label definitions augmented interaction model for legal charge prediction, in: Advances in Information Retrieval: 43rd European Conference on IR Research, ECIR 2021, Virtual Event, March 28 –April 1, 2021, Proceedings, Part I, Springer-Verlag, Berlin, Heidelberg, 2021, p. 270–283. URL: https://doi.org/10.1007/ 978-3-030-72113-8_18. doi:10.1007/978- 3- 030- 72113- 8_18. [17] B. Luo, Y. Feng, J. Xu, X. Zhang, D. Zhao, Learning to predict charges for criminal cases with legal basis, in: Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Copenhagen, Denmark, 2017, pp. 2727–2736. URL: https://aclanthology.org/D17-1289. doi:10.18653/v1/ D17- 1289. [18] L. Gan, K. Kuang, Y. Yang, F. Wu, Judgment prediction via injecting legal knowledge into neural networks, Proceedings of the AAAI Conference on Artificial Intelligence 35 (2021) 12866–12874. URL: https://ojs.aaai.org/index.php/AAAI/article/view/17522. [19] W. Yang, W. Jia, X. Zhou, Y. Luo, Legal judgment prediction via multi-perspective bifeedback network, in: Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19, International Joint Conferences on Artificial Intelligence Organization, 2019, pp. 4085–4091. URL: https://doi.org/10.24963/ijcai.2019/567. doi:10.24963/ijcai.2019/567. [20] N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Gesmundo, M. Attariyan, S. Gelly, Parameter-eficient transfer learning for nlp, in: International Conference on Machine Learning, PMLR, 2019, pp. 2790–2799. [21] X. L. Li, P. Liang, Prefix-tuning: Optimizing continuous prompts for generation, in: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), Association for Computational Linguistics, Online, 2021, pp. 4582–4597. URL: https://aclanthology.org/2021.acl-long.353. doi:10.18653/v1/2021.acl- long.353. [22] J. He, C. Zhou, X. Ma, T. Berg-Kirkpatrick, G. Neubig, Towards a unified view of parameter-eficient transfer learning, in: ICLR, 2022. [23] H. Liu, D. Tam, M. Muqeeth, J. Mohta, T. Huang, M. Bansal, C. A. Rafel, Few-shot parameter-eficient fine-tuning is better and cheaper than in-context learning, Advances in Neural Information Processing Systems 35 (2022) 1950–1965. [24] T. Wolf, L. Debut, V. Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P. von Platen, C. Ma, Y. Jernite, J. Plu, C. Xu, T. L. Scao, S. Gugger, M. Drame, Q. Lhoest, A. M. Rush, Transformers: State-of-the-art natural language processing, in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, Association for Computational Linguistics, Online, 2020, pp. 38–45. URL: https://www.aclweb.org/anthology/2020. emnlp-demos.6. [25] S. Mangrulkar, S. Gugger, L. Debut, Y. Belkada, S. Paul, Peft: State-of-the-art parametereficient fine-tuning methods, https://github.com/huggingface/peft, 2022.