CorDA: Context-Oriented Decomposition Adaptation of Large Language Models for Task-Aware Parameter-Efficient Fine-tuning

Current parameter-efficient fine-tuning (PEFT) methods build adapters widely agnostic of the context of downstream task to learn, or the context of important knowledge to maintain. As a result, there is often a performance gap compared to full-parameter fine-tuning, and meanwhile the fine-tuned model suffers from catastrophic forgetting of the pre-trained world knowledge. In this paper, we propose CorDA, a Context-oriented Decomposition Adaptation method that builds learnable task-aware adapters from weight decomposition oriented by the context of downstream task or the world knowledge to maintain. Concretely, we collect a few data samples, and perform singular value decomposition for each linear layer of a pre-trained LLM multiplied by the covariance matrix of the input activation using these samples. The inverse of the covariance matrix is multiplied with the decomposed components to reconstruct the original weights. By doing so, the context of the representative samples is captured through deciding the factorizing orientation. Our method enables two options, the knowledge-preserved adaptation and the instruction-previewed adaptation. For the former, we use question-answering samples to obtain the covariance matrices, and use the decomposed components with the smallest $r$ singular values to initialize a learnable adapter, with the others frozen such that the world knowledge is better preserved. For the latter, we use the instruction data from the fine-tuning task, such as math or coding, to orientate the decomposition and train the largest $r$ components that most correspond to the task to learn. We conduct extensive experiments on Math, Code, and Instruction Following tasks.

Paper

Similar papers

Peer review

Reviewer kduk6/10 · confidence 4/52024-06-22

Summary

This paper proposes CorDA, a context-oriented decomposition adaptation method that initializes LoRA adapter with different components from the weight decomposition to support two different options: knowledge-preserved adaptation and instruction-previewed adaptions. Through extensive experiments on LLaMA-2-7b and RoBERTa, the paper shows that the knowledge-preserved adaptation can better preserve performance on general knowledge task and the instruction-previewed adaptations can yield better target task performance compared to the baseline PEFT methods.

Strengths

1. The experimental results on the instruction-previewed mode are strong. Table 2 and Table 3 demonstrate that the purposed method yields better performance when fine-tuning on targets compared with other PEFT methods and its performance is nearly on par with full finetuning. 2. The proposed method is intuitive and simple in concept. If the decoupled components capture context in the principle components decoupled, the two modes are well-motivated.

Weaknesses

1. One major concern of this work is the lack of theoretical support, and some arguments in the paper are not well justified. - Line 45-47: “The covariance matrix of each layer’s activation … are responsive to the task triggered to highlight different aspects of the pre-train weight.” Is there any support for this or is this a pure intuition? - Why do you perform SVD on $WC$ ? What does $WC$ represent? 2. The knowledge-preserved mode is not very compelling given this method will hurt the target task performance and cannot achieve the best of both worlds by design. Although the authors have already recognized this when stating the limitations, I don’t know why the paper emphasizes this mode in parallel with the instruction-previewed adaptation mode. Regarding the design of method to achieve the best of both worlds, [1] may be a pointer for reference as it also studies how to preserve the forgetting of general knowledge in LMs when continual training them. [1] Continual Pre-training of Language Models, Ke et al., ICLR 2023 **Update:** Most of points are addressed by the author response, so I update my score from 4 to 6.

Questions

1. Does the major difference between CorDA and LoRA lie in the initialization? Have you compared with other PEFT methods that focuses on initialization technique? Currently, the baseline comparison is not very comprehensive.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes, it's discussed from Line 293.

Authorsrebuttal2024-08-07

Discussion about the value of knowledge-preserved mode and answers to remaining questions

> (2) The knowledge-preserved mode is not very compelling given this method will hurt the target task performance. I don’t know why the paper emphasizes this mode in parallel with the instruction-previewed adaptation mode. [1] may be a pointer for reference as it also studies how to preserve the forgetting of general knowledge in LMs when continual training them. **Please note that the knowledge-preserved adaptation does NOT hurt the target task performance**. As shown in Table 1, compared with LoRA, our method has better performance than LoRA in both worlds (world knowledge benchmarks and downstream tasks) in most cases. We **quote some results of LoRA and CorDA in Table 1 here** for your reference. |Method|Trivia QA | NQ open | WebQS | GSM8k | Math | Avg. | |---|---|---|---|---|---|---| |LoRA|44.17 | 1.91 |6.64| 42.68| 5.92| 20.26| |CorDA | 44.30 | 9.36 | 7.14 | 44.58 | 6.92 | 22.46 | |Method|Trivia QA | NQ open | WebQS | MTBench | Avg.| |---|---|---|---|---|---| |LoRA|47.46 |10.28| 7.73| 4.60| 17.52| |CorDA | 50.34 |14.43| 8.17| 5.05| 19.50| More importantly, as we respond above, existing PEFT methods rarely consider or support finetuning with knowledge better perserved. There are some studies on the continual training of LLMs [1] and [14, 21] (ref. in our paper), but they are not PEFT methods. We will cite [1] in the revised version of our paper. As shown in Table 1, when using the average performance of the both worlds to measure the comprehensive ability, CorDA in knowledge-preserved adaptation achieves the best results among all the compared methods in all the three tasks. **Therefore, we respectfully do not agree with your comment that "knowledge-preserved adaptation is not compelling" and "it hurts the target task performance".** Admittedly, CorDA in knowledge-preserved adaptation is not stronger than the instruction-previewed adaptation when only evaluating the downstream task performance. But please note that maintaining pre-trained knowledge better and pursuing a better finetuning performance is inherently a tradeoff (also mentioned in [14] and [21]). A method may be better than another one in both worlds, just like CorDA v.s. LoRA in the results above, but for the method itself, the two worlds are still a tradeoff. We would like to use an analogy to support. The advanced neural architecture (_e.g._ Transformer) may be better than a traditional MLP/CNN architecture in terms of both accuracy and parameter efficiency. But for the better architecture itself, higher accuracy still brings more parameters (BERT_large is better than BERT_base with more parameters). Therefore, when knowledge maintenance is not a concern, we introduce our instruction-previewed adaptation, which puts all its efforts for the downstream task, surpassing the competitive studies DoRA and PiSSA in the three finetuning tasks, Math, Code, and instruction following, as shown in Table 2. In conclusion, the knowledge-preserved adaptation and instruction-previewed adaptation highlight the comprehensive performance and the specialized ability of downstream task, respectively. We think this is also a feature of our PEFT method, allowing for customized selectivity based on the actual need. That is why we emphasize the two modes in parallel. [1] Continual Pre-training of Language Models, Ke et al., ICLR 2023 > (3) Questions: Does the major difference between CorDA and LoRA lie in the initialization? Have you compared with other PEFT methods that focuses on initialization technique? Currently, the baseline comparison is not very comprehensive. Yes, CorDA brings task context into the LoRA adapter initialization. Adopting the same LoRA structure not only facilitates fair comparison, but also enables to restore the orignal LLM architecture after finetuning without architectural change or introducing inference burden. Yes, we have compared with PiSSA, which also focues on the LoRA adapter initialization but does not consider task context. DoRA builds the adapter with a normalization and a learnable magnitude, and also does not consider task context. It is noteworthy that both DoRA (ICML 24) and PiSSA (released on Arxiv in April 2024) are recent studies and are strong baselines. Besides, full parameter finetuning is the most direct reference because it usually has the best finetuning performance without considering parameter efficiency. For downstream tasks, as shown in Table 2 and Table 3, our method achieves finetuning performances on par with full parameter finetuning, and better performances than the compared PEFT methods LoRA, DoRA, and PiSSA. For comprehensive ability (with knowledge benchmarks included), as shown in Table 1, our method has the best average performance among full finetuning and the PEFT methods. Therefore, our experimental results are already able to demonstrate the effectiveness of our proposed method.

Reviewer kduk2024-08-10

I thank the authors for the detailed response. The first part of the response adequately addresses my Weakness 1. However, it's worth noticing that I am not saying every paper necessarily needs to have a theoretical part. Solid, inspiring empirical results are very important for AI/ML application. The reason why I raised Weakness 1 is the current writing of Section 3.2 gives people an impression that the paper states that the covariance matrix is tightly connected with task context/patterns, where the latter concept itself does not have a formal definition (it's really hard to articulate what task context is). I suggest weakening the argument here and discussing the empirical comparison more, as it's the empirical results that show decomposing $WC$ could lead to a better results. Overall, the author response addresses my major concern about the method proposed in this paper, so I increase my score accordingly.

Authorsrebuttal2024-08-10

Dear Reviewer kduk, Thank you for increasing the score and your suggestions. We will rephrase the argument, and include more discussion about the empirical comparison in the revised paper. Authors

Reviewer HCBt7/10 · confidence 3/52024-07-09

Summary

The paper proposes a context-oriented decomposition adaptation method for large language models (LLMs) called CorDA. This method constructs learnable adapters by considering the context of downstream tasks or world knowledge. It can bridge the performance gap between parameter-efficient fine-tuning (PEFT) and full-parameter fine-tuning, while also mitigate the problem of catastrophic forgetting. Specifically, the method decomposes the weights of LLMs using singular value decomposition (SVD) guided by the covariance matrix of input activations. Through this approach, the model can identify important weights related to specific contexts, thereby creating context-aware adapters. These adapters can be customized for different tasks, enabling the model to retain general knowledge while enhancing performance on specific tasks.

Strengths

S1: The structure of the paper is relatively clear. The introduction progressively leads readers to understand the current state of fine-tuning LLMs and the challenges posed by knowledge forgetting after fine-tuning. This approach allows readers to quickly grasp the current situation of LLMs fine-tuning. In the subsequent method section, related modules are introduced based on these challenges. S2:The main challenge addressed in this paper is the catastrophic forgetting of knowledge when fine-tuning LLMs. To tackle this, the authors propose a context-oriented decomposition method and introduce two modules. The first module generates an adapter that retains the model’s general knowledge by activating the covariance matrix using other question-answering datasets. The second module, designed to adapt to task-specific instructions, is called the instruction preview adaptation module, which generates adapters specific to the given tasks. The paper thoroughly explains the implementation process of the method, including the calculation of the covariance matrix and the application of SVD. S3:The experimental validation in this paper is comprehensive, verifying the effectiveness of each module. The results of the modules are thoroughly analyzed through extensive experiments, demonstrating the effectiveness of the proposed modules.

Weaknesses

S1:In the experimental section, it is necessary to add a description of the experimental environment and provide parameters to offer readers more possibilities for replication and ensure authenticity. Additionally, the paper lacks specific quantification of catastrophic forgetting in the model.

Questions

none

Rating

7

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

none

Reviewer Mgdn6/10 · confidence 3/52024-07-13

Summary

This paper introduces a new parameter-efficient fine-tuning (peft) method called Context-oriented Decomposition Adaptation (CorDA). Although fundamentally similar to LoRA, it differs in that it initializes two low-rank matrices using the SVD results of pre-trained weights reflecting the data context. To incorporate the data context, few samples specific to the task are required, utilizing the input activations of these samples for the large language model (LLM). CorDA can be applied in two ways: knowledge-preserved adaptation and instruction-previewed adaptation. The former aims to retain specific knowledge during fine-tuning, while the latter is used to adapt the LLM more effectively to the desired instructions.

Strengths

- **Originality**: The approach of modifying the LoRA method to maintain or enhance the performance of language models on specific data types appears to be a novel attempt. This method has not been explored in existing works dealing with parameter-efficient fine-tuning, thus offering originality. - **Quality**: Experimental results demonstrate that CorDA is more effective than existing methods like LoRA or PiSSA in fine-tuning models such as Llama and RoBERTa. - **Significance**: The idea of using data to derive effective low-rank matrices could inspire future research.

Weaknesses

- **Originality**: There are methodological similarities with PiSSA, and comparisons with existing works like AdaLoRA [1] are lacking. - **Quality**: The exact role of the covariance matrix $C$ is not clearly explained. The paper states in lines 145-147 that “the covariance matrix of each layer’s activation will exhibit different outlier patterns as they are responsive to the task triggered to highlight **different aspects** of the pre-trained weight.” Providing visualizations to show how data from different tasks trigger different parts of the weight would enhance clarity. - **Clarity**: The amount and type of data used to derive the covariance matrix $C$ significantly impact performance, but the paper lacks a clear discussion and analysis on this. Moreover, there is no guidance on how to choose $r$ in Equations 4 and 5, which appears to be a crucial hyperparameter. This omission makes it challenging to practically apply the proposed method based solely on the information provided in the paper. - **Significance**: Although the research seems advantageous compared to existing parameter-efficient fine-tuning methods, it is unclear how much and what quality of data is needed to derive $C$ for better performance. The usability of the CorDA heavily depends on the amount and quality of data required. Despite this, the concept of using data to derive better low-rank matrices is novel. Reference [1] AdaLoRA: Adaptive Budget Allocation For Parameter-efficient Fine-tuning

Questions

- It would be beneficial to include visualizations of the covariance matrix derived from different data. - The paper should incorporate a data selection strategy for $C$ and an extensive analysis on $r$.

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

The authors adequately addressed the limitations in Section 4.4.

Reviewer 527n7/10 · confidence 4/52024-07-14

Summary

The paper proposes an initializes algorithm for LoRA based fine-tuning, with the aim to maintain world knowledge and also improve training performance on the fine-tuning task at hand. The authors carefully explain the reasoning behind their initialization scheme, and conduct extensive studies to showcase the efficacy of their framework. Furthermore, they show that they can get surprising improvements on MT bench, simply by maintaining knowledge information. Overall, the paper presents a systematic approach to initializing LoRA parameters, and the general principle can guide the community towards better training of large language models.

Strengths

The strength of the paper lies in its simplistic exposition of its motivation, proposed framework, and the extensive experimental study to show the efficacy of the initialization framework. The authors begin with a clean exposition to SVD, and how one can think of the different eigenvectors in the input weight covariance matrix. With a very intuitive discussion, the authors show the directions along which the weight parameters aren't perturbed during training to maintain world knowledge. The directions are selected by looking at the covariance matrix of knowledge-test datasets like TriviaQA, and NaturalQA. Furthermore, the authors indicate the directions along which they want to pre-capture features of the fine-tuning task at hand, which can give bigger returns compared to training LoRA parameters from scratch. Overall, the principle of data-dependent initialization is novel and can guide the community towards more systematic training designs.

Weaknesses

Overall, I don't see much weakness with the work. I have a few questions regarding the experimental framework, which I would like to discuss during the rebuttal period. - How did the authors decide to maintain the last $r$ eigenvectors for world-knowledge? I believe, the authors should cite relevant works or make proper justification on this design choice. - How scalable is their proposed approach to other datasets where we may want to maintain knowledge? That is, how likely is the model to maintain knowledge on TriviaQA, if it wasn't included during LoRA initialization? - What is the necessary sample complexity for initializing the LoRA parameters, i.e. how many samples from the knowledge datasets are necessary to get a good estimate of the directions to freeze? - Is the knowledge preservation necessary at each weight parameters and each layer of the model? How do results change when the model's LoRA parameters are restricted only in the lower layers, while the higher layers are given more freedom during fine-tuning?

Questions

Please check my questions above.

Rating

7

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

The authors discuss the limitations of their work in section 4.4, and clearly indicate the future directions that the community can pursue starting from their work.

Reviewer Mgdn2024-08-11

Thank the author for the clarifications. The response effectively addresses my concerns, clearly highlighting the key contributions of this work in comparison to PiSSA, as well as the different patterns of covariance matrices across tasks. Therefore, I have increased the score accordingly.

Reviewer 527n2024-08-12

I thank the authors for their response. After going through the response, I am maintaining my score.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC