MFTCoder: Boosting Code LLMs with Multitask Fine-Tuning

Abstract: Code LLMs have emerged as a specialized research field, with remarkable studies dedicated to enhancing model's coding capabilities through fine-tuning on pre-trained models. Previous fine-tuning approaches were typically tailored to specific downstream tasks or scenarios, which meant separate fine-tuning for each task, requiring extensive training resources and posing challenges in terms of deployment and maintenance. Furthermore, these approaches failed to leverage the inherent interconnectedness among different code-related tasks. To overcome these limitations, we present a multi-task fine-tuning framework, MFTcoder, that enables simultaneous and parallel fine-tuning on multiple tasks. By incorporating various loss functions, we effectively address common challenges in multi-task learning, such as data imbalance, varying difficulty levels, and inconsistent convergence speeds. Extensive experiments have conclusively demonstrated that our multi-task fine-tuning approach outperforms both individual fine-tuning on single tasks and fine-tuning on a mixed ensemble of tasks. Moreover, MFTcoder offers efficient training capabilities, including efficient data tokenization modes and PEFT fine-tuning, resulting in significantly improved speed compared to traditional fine-tuning methods. MFTcoder seamlessly integrates with several mainstream open-source LLMs, such as CodeLLama and Qwen. Leveraging the CodeLLama foundation, our MFTcoder fine-tuned model, \textsc{CodeFuse-CodeLLama-34B}, achieves an impressive pass@1 score of 74.4\% on the HumaneEval benchmark, surpassing GPT-4 performance (67\%, zero-shot). MFTCoder is open-sourced at \url{https://github.com/codefuse-ai/MFTCOder}

• The paper presents a multitask fine-tuning framework that significantly enhances code LLM performance by integrating balanced loss functions and PEFT techniques.
• The methodology leverages dynamic padding, self-instruct dataset creation, and LoRA/QLoRA to reduce computational costs and improve training efficiency.
• Experimental results show robust gains across 5 code-related tasks, with MFT-trained models outperforming individually fine-tuned counterparts and GPT-4 benchmarks.

Analysis of "MFTCoder: Boosting Code LLMs with Multitask Fine-Tuning"

The paper "MFTCoder: Boosting Code LLMs with Multitask Fine-Tuning" presents a comprehensive exploration of a multitask fine-tuning framework for LLMs in the domain of code generation. The work addresses the limitations of traditional fine-tuning approaches, which often require separate tuning for each task, incurring substantial computational costs and failing to exploit the interconnectedness of code-related tasks. The authors propose MFTCoder, a framework that enables concurrent fine-tuning across multiple tasks, ensuring efficient deployment and enhanced performance.

Methodology

MFTCoder leverages a multitask fine-tuning approach, incorporating various loss functions to handle issues such as data imbalance, differing task difficulties, and convergence speed disparities. The method effectively combines multiple tasks, providing a unified training framework that is both resource-efficient and performance-oriented.

The authors employed several techniques to enhance MFTCoder's efficiency:

1. Instruction Dataset Construction: Utilizing self-instruct methodologies and agent-driven conversations to automate the creation of training datasets.
2. Efficient Tokenization Modes: Implementing dynamic padding and packing strategies to minimize padding token overhead and improve training speed.
3. Parameter-Efficient Fine-Tuning (PEFT): Adopting LoRA and QLoRA techniques to reduce the parameter load and enable fine-tuning of extensive models on limited resources.
4. Balanced Loss Functions: Designing varying loss functions to ensure balanced training among tasks, specifically addressing data imbalances and convergence issues.

Experimental Evaluation

The efficacy of MFTCoder was demonstrated through extensive experiments on 7 different models across 5 diverse code-related tasks, including code completion, text-to-code generation, code comment generation, code translation, and unit test generation. Key findings include:

• MFTCoder outperformed individually fine-tuned models (SFT-S) and models fine-tuned on mixed tasks (SFT-Mixed).
• Notably, the MFT-trained model demonstrated superior generalization capabilities on unseen tasks, such as text-to-SQL generation.
• On the HumanEval benchmark, MFTCoder's CodeFuse-CodeLLama-34B achieved a pass@1 score of 74.4%, exceeding the performance of GPT-4 (67%, zero-shot).

Implications and Future Directions

The results illustrate that multitask fine-tuning using MFTCoder significantly boosts the performance of code LLMs, offering a scalable and efficient solution for handling multiple code-oriented tasks. This work has profound implications for both theoretical research and practical applications in AI, suggesting that multitask frameworks could usher in more flexible, robust, and comprehensive AI systems.

Future developments may focus on refining task division strategies, exploring more adaptive multitask optimization techniques, and extending the framework to broader NLP and AI domains. Additionally, further investigation into the balancing of convergence speeds across diverse tasks could enhance adaptability and efficiency.

In summary, this paper lays a solid foundation for improved multitask learning practices in AI, providing a detailed roadmap for enhancing code LLM capabilities through finely-tuned, comprehensive training regimes.