Star-Agents: Automatic Data Optimization with LLM Agents for Instruction Tuning

The efficacy of large language models (LLMs) on downstream tasks usually hinges on instruction tuning, which relies critically on the quality of training data. Unfortunately, collecting high-quality and diverse data is both expensive and time-consuming. To mitigate this issue, we propose a novel Star-Agents framework, which automates the enhancement of data quality across datasets through multi-agent collaboration and assessment. The framework adopts a three-pronged strategy. It initially generates diverse instruction data with multiple LLM agents through a bespoke sampling method. Subsequently, the generated data undergo a rigorous evaluation using a dual-model method that assesses both difficulty and quality. Finaly, the above process evolves in a dynamic refinement phase, where more effective LLMs are prioritized, enhancing the overall data quality. Our empirical studies, including instruction tuning experiments with models such as Pythia and LLaMA, demonstrate the effectiveness of the proposed framework. Optimized datasets have achieved substantial improvements, with an average increase of 12% and notable gains in specific metrics, such as a 40% improvement in Fermi, as evidenced by benchmarks like MT-bench, Vicuna bench, and WizardLM testset.

Paper

References (47)

Scroll for more · 35 remaining

Similar papers

Peer review

Reviewer 4rks6/10 · confidence 3/52024-07-10

Summary

The paper introduces the Star-Agents framework, an innovative system designed to enhance the quality of datasets used for instruction-tuning of large language models (LLMs). The framework addresses the challenges of collecting high-quality and diverse data by automating the process through multi-agent collaboration and assessment. The optimized datasets resulted in significant performance improvements. The research contributes to an advanced automated data optimization system that refines instruction samples with suitable complexity and diversity for LLMs, leading to more efficient model alignment and improved performance.

Strengths

1. The paper presents a novel approach to enhancing the quality of datasets for instruction-tuning of LLMs. The concept of using multiple agents to generate diverse instruction data and the dual-model evaluation strategy for data selection are innovative. 2. The paper includes comprehensive experiments with different models, showcasing the effectiveness of the Star-Agents framework. 3. The reported performance improvements of up to 12% on average, and over 40% in specific metrics, highlight the practical significance of the research.

Weaknesses

1. This paper utilizes multi-agent collaboration, but it seems that the paper does not explore the impact of different numbers of agents or agent pairs on the results. The authors should also specify the exact number of agents used in the experiments in the "Implementation Details" section.

Questions

1. Could you provide some specific cases? For example, demonstrate how the outputs from different agent pairs differ. 2. What impact would removing some or adding additional agent pairs have on the results?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have mentioned the limitations of this work in the paper and have also suggested potential research directions for addressing these limitations.

Authorsrebuttal2024-08-11

Thanks for the comments.

Dear Reviewer 4rks, We sincerely appreciate the time you’ve taken to provide valuable feedback on our paper. In our rebuttal, we have thoroughly addressed all of your initial concerns and included the requested experimental results. If you have any further questions or concerns, we would be happy to discuss them with you. Additionally, we welcome any new suggestions or comments you may have!

Reviewer b1DV6/10 · confidence 4/52024-07-12

Summary

The paper introduces the "Star-Agents" framework, designed to optimize data for instruction tuning in large language models (LLMs). This system automates the process of enhancing dataset quality by employing a multi-agent approach. Empirical studies demonstrate that the optimized datasets lead to significant performance improvements in models such as Pythia and LLaMA.

Strengths

1. The motivation is clear and it is easy to follow their method. 2. The framework's effectiveness is validated through extensive experiments, showing significant performance improvements in various benchmarks.

Weaknesses

1. What is the overhead of this proposed method, like wall-clock time? 2. stability: can this method be scalable to large scale dataset optimization, like web-crawl data?

Questions

Please refer to the above.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Authorsrebuttal2024-08-11

Thanks for the comments.

Dear Reviewer b1DV, We sincerely appreciate the time you took to provide valuable comments on our paper. In our rebuttal, we have addressed all of your initial concerns and included the requested experimental results. If you still have any question or concern, we would be glad to discuss them with you. Additionally, we welcome any new suggestions or comments you may have! Regards

Reviewer kHqZ6/10 · confidence 3/52024-07-15

Summary

The Star-Agents framework presents an advanced approach for enhancing data quality in instruction-tuning of large language models (LLMs) through multi-agent collaboration and automated assessment. By generating diverse instruction data using various LLM agents and evaluating it with a Dual-model metric, this framework ensures both diversity and quality. It dynamically refines its processes to prioritize more effective agents, leading to significant performance improvements. Experimental results show an average increase of 12% in model performance, with some metrics seeing increases up to 40%. This method addresses common issues in data generation such as lack of stylistic variety and overly complex datasets, making it a robust solution for optimizing LLM training datasets.

Strengths

The paper stands out for its innovative dual-model scoring system, which refines data quality assessment for large language models (LLMs) and is a key contribution to the field. It demonstrates state-of-the-art results, with the framework achieving an average performance increase of 12% and up to 40% in specific metrics, highlighting its effectiveness. Additionally, the paper is clearly written and well-structured, making complex concepts accessible and easy to follow, which enhances its academic impact and usability.

Weaknesses

I could not find any weakness apart from mentioned in the questions below.

Questions

What is the size and composition of the datasets used in your experiments? Have you compared your framework's performance using standard benchmarks like MMLU or those from Hugging Face? What would be the impact of using only Mistral-ChatGPT for 70,000 iterations on the diversity and quality of the generated data?

Rating

6

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

The authors acknowledge a limitation in their framework, noting that it is currently designed for optimizing single-turn interactions. They suggest that extending this approach to multi-turn scenarios could further enhance its applicability and effectiveness, addressing more complex dialogue systems where contextual continuity and conversational depth are critical.

Authorsrebuttal2024-08-11

Thanks for your feedback!

Dear Reviewer kHqZ, Thank you for your thoughtful feedback and valuable comments. We appreciate the opportunity to further clarify our methodology. Regarding the metric used to calculate diversity, we employed cosine similarity as the primary metric. Specifically, we calculated the cosine similarity between the sentence embeddings of the datasets. The diversity metric is inversely related to the cosine similarity; hence, lower cosine similarity indicates higher diversity in the generated data. In our analysis, we found that the Mistral-ChatGPT dataset exhibited a cosine similarity of approximately 20.8%, which indeed suggests a higher degree of similarity (and thus lower diversity) when compared to the dataset generated using all agent pairs, which had a cosine similarity of 6.9%. We hope this clarification adequately addresses your concerns, and we are grateful for your guidance in improving the clarity and rigor of our work.

Reviewer piXU5/10 · confidence 4/52024-07-17

Summary

The paper presents a framework for enhancing the quality of instruction datasets used for tuning large language models (LLMs). The proposed framework, Star-Agents, leverages multi-agent collaboration to generate, evaluate, and refine instruction data automatically. The approach comprises three main components: 1. **Diverse Data Generation**: The framework employs multiple LLM agents to generate varied instruction data. Each agent pair, consisting of distinct instruction and response agents, creates diverse data samples to enrich the dataset. 2. **Dual-model Evaluation**: This component introduces a two-tiered evaluation mechanism using both small and large models. The evaluation metric considers the difficulty and quality of the generated data, ensuring it is challenging yet manageable for the target model. 3. **Dynamic Refinement**: The framework dynamically adjusts the sampling probabilities of agent pairs based on their performance, promoting the selection of the most effective agents for data generation. Empirical studies demonstrate the optimized datasets generated by Star-Agents led to performance improvements, with an average increase of 12% and up to 40% in specific benchmarks such as MT-bench, Vicuna-bench, and the WizardLM testset. ------------------------------------------------------------------------------------------------------------------------------------------------------------ Thank you for your reply and I have updated my score accordingly.

Strengths

**Originality** The paper introduces a framework, Star-Agents, for automatic data optimization in instruction tuning for large language models (LLMs). This originality arises from: - **Multi-Agent Collaboration**: Utilizing multiple LLM agents to generate diverse and high-quality instruction data. This approach addresses the common limitations of single-model data generation methods, ensuring a richer and more varied dataset. - **Dual-Model Evaluation**: Implementing a dual-model evaluation strategy that assesses both the difficulty and quality of the generated data. This innovative metric ensures the data is challenging yet manageable, enhancing the instruction tuning process. - **Dynamic Refinement**: The dynamic adjustment of sampling probabilities for agent pairs based on performance is a creative mechanism that optimizes the data generation process over time. **Quality** The paper is supported by empirical studies. The framework is tested using instruction-tuning experiments with two models, including Pythia and LLaMA. The results consistently show performance improvements, validating the quality of the proposed method. **Clarity** The paper is clearly written and well-structured, facilitating easy understanding of the proposed framework.

Weaknesses

**Limited Dataset Evaluation** The paper evaluates the Star-Agents framework on a limited set of datasets, which may not fully capture the framework's robustness and generalizability. Specifically: - **Small Evaluation Datasets**: The datasets used for evaluations are relatively small. Evaluating the framework on larger, more diverse datasets would provide a better understanding of its effectiveness across different data scales and domains. **Complexity of the Method** The Star-Agents framework involves multiple stages and the use of several teacher models, which introduces complexity: - **Multiple Teacher Models**: The requirement for multiple LLM agents increases the computational cost and complexity of the implementation. Simplifying the framework or exploring methods to reduce the number of required teacher models could enhance its practical applicability. - **Three-Stage Process**: The three-pronged approach, while thorough, may be overly complex for some applications. Streamlining the process without sacrificing performance gains could make the framework more accessible and easier to implement. **Lack of Human Evaluations** The paper does not include human evaluations of test data, which is crucial for assessing the practical quality and usability of the model. "Absence of Standard Benchmark Results" The framework's performance is not evaluated on widely recognized benchmarks, limiting the comparability of its results: - **Standard Benchmarks**: Including evaluations on standard benchmarks such as GSM8K, HumanEval, or other well-known datasets would provide a clearer comparison with existing methods. This would help situate the Star-Agents framework within the broader context of instruction tuning and LLM performance.

Questions

What will be the results if base models are larger, e.g. llama 13b or llama 70b?

Rating

5

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

See more details in Weaknesses section.

Authorsrebuttal2024-08-11

Thanks for the comments.

Dear Reviewer piXU, We sincerely appreciate the time and effort you invested in providing valuable feedback on our paper. In our rebuttal, we have thoroughly addressed all of your initial concerns and included the requested experimental results. If you have any further questions or concerns, we would be happy to discuss them with you. Additionally, we welcome any new suggestions or comments you might have! Regards

Reviewer kHqZ2024-08-09

Thank you for addressing most of my concerns. I have few comments. "Thank you for your insightful question regarding the impact of using Mistral-ChatGPT for 70,000 iterations on the diversity and quality of the generated data. We have conducted a detailed analysis to address your concerns. Firstly, we generated 70,000 data samples exclusively using the agent-pair Mistral-ChatGPT. In terms of diversity, we observed a decline in the variation of responses for the same instructions. Regarding quality, some of the data generated was of lower quality compared to using the all of agent-pairs, but it was still superior to the baseline data. To further assess the implications, we trained the model using this generated dataset. The performance metrics are presented in the Table 2. The results demonstrate that relying on Mistral-ChatGPT for data generation negatively impacts the model's overall capability compared to using the full set of agent-pairs, but it is still superior to the baseline. This confirms the effectiveness of our approach." Here what metric did you use to calculate diversity?

Reviewer 4rks2024-08-12

Response to the authors

Thanks to the authors for resolving my concern. For the Q1, from the table, it seems that 10 agent pairs is not the number that achieve the top of performance. What still concerns me is the balence of performance and cost (if more agent pairs cost more).

Authorsrebuttal2024-08-13

Thanks for your feedback!

Dear Reviewer 4rks, Thank you for your valuable feedback. We understand your concern regarding the balance between performance and cost when increasing the number of agent pairs. To address this, we conducted additional experiments. As shown in Table 2, we observed that the performance improvement starts to plateau when the number of agent pairs increases to 8. While the model's performance is highest with 10 agent pairs, the difference between 8 and 10 pairs is marginal. Considering the trade-off between performance and cost, we believe that using 8–10 agent pairs is a reasonable and effective choice. This range provides a good balance, ensuring high performance while keeping the cost manageable. Table 2. Performance using different agent pairs | model | agent-pair | Vicuna-bench | Wizardlm-bench | MT-bench | Average | | ------------------------------- | ---------- | ------------ | -------------- | -------- | ------- | | Pythia-1B-evol_instruct | 1 | 5.07 | 3.55 | 2.56 | 3.73 | | Pythia-1B-star_instruct_2_pair | 2 | 5.23 | 3.88 | 2.69 | 3.93 | | Pythia-1B-star_instruct_5_pair | 5 | 5.29 | 3.87 | 2.78 | 3.98 | | Pythia-1B-star_instruct_7_pair | 7 | 5.48 | 3.92 | 2.79 | 4.06 | | Pythia-1B-star_instruct_8_pair | 8 | 5.69 | 3.92 | 2.78 | 4.13 | | Pythia-1B-star_instruct_9_pair | 9 | 5.80 | 3.91 | 2.77 | 4.16 | | Pythia-1B-star_instruct_10_pair | 10 | 5.93 | 3.9 | 2.69 | 4.17 |

Reviewer b1DV2024-08-13

Thanks for your response and it would be great if we can add this in the camera ready version. I will keep my score.

Authorsrebuttal2024-08-14

Thanks for your feedback!

Dear Reviewer b1DV, Thanks for your feedback and valuable suggestions! We will ensure that the recommended additions are included in the camera-ready version of the paper. Regards

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC