alexshengzhili/mllm-dpo
收藏资源简介:
--- license: apache-2.0 --- [[Arxiv paper](https://arxiv.org/abs/2402.10884)] [[GitHub](https://github.com/findalexli/mllm-dpo)] [[Data](https://huggingface.co/datasets/alexshengzhili/llava-preference-dpo)] [[Model](https://huggingface.co/alexshengzhili/llava-v1.5-13b-dpo/edit/main/README.md)] [[Data](https://huggingface.co/datasets/alexshengzhili/mllm-dpo)] **Release Arxiv:** [https://arxiv.org/abs/2402.10884](https://arxiv.org/abs/2402.10884) **Developers:** Shengzhi Li (TIFIN), Rongyu Lin (KAUST), Shichao Pei (University of Massachusetts Boston) **Affiliations:** TIFIN, KAUST, University of Massachusetts Boston **Contact Information:** alex.li@tifin.com, rongyu.lin@kaust.edu.sa, shichao.pei@umb.edu ## Overview The llava-v1.5-13b-dpo model and corresponding data set is designed to enhance the instruction-following capabilities of multi-modal large language models (MLLMs), particularly in scenarios where visual instruction tuning might degrade language proficiency. This model leverages a novel Direct Preference Optimization (DPO) method, along with a curated 6K-entry VQA preference dataset, to achieve superior performance on multi-modal tasks and benchmarks. ## Intended Use - **Primary Applications:** This model is intended for tasks requiring the integration of text and image modalities, including but not limited to visual question answering (VQA), image captioning, and multi-modal instruction following. - **Target Audience:** Researchers and practitioners in the fields of natural language processing, computer vision, and multi-modal AI. ## Training Data The MM-LLM-DPO model was trained using a lightweight (6k entries) VQA preference dataset, where answers were annotated for 5 quality metrics in a granular fashion. The dataset was designed to address the diversity and complexity gap typically observed in VQA datasets. ## Evaluation The model demonstrates significant improvements over baseline models like Vicuna and LLaVA on various benchmarks: - **MT-Bench:** Achieved a score of 6.73, surpassing Vicuna's 6.57 and LLaVA's 5.99. - **Visual Instruction Performance:** Recorded a +4.9% improvement on MM-Vet and +6% on LLaVA-Bench. | Model Name | MM-Vet | LLaVA-bench | PoPe | MM-Bench | MT-bench | AlpacaEval | |------------|--------|-------------|-------|----------|----------|------------| | **Vicuna-1.5-13b [16]** | - | - | - | - | 6.57 | 81.4 | | **LLaVA-1.5-13b [10]** | 36.3 | 73.1 | 0.859 | 67.4 | 5.99 | 79.3 | | **LLaVA-RLHF-13b [23]**| 37.2 | 76.8 | 0.869 | 60.1 | 6.18 | 81.0 | | **Standard SFT** | 36.5 | 63.7 | 0.850 | 65.4 | 5.01 | 50.2 | | **SteerLM** | 35.2 | 67.0 | 0.878 | 65.1 | 5.70 | 68.8 | | **Rejection-sampling** | 38.0 | 70.6 | 0.883 | 67.6 | 6.22 | 74.9 | | **llava-v1.5-13b-dpo** | 41.2 | 79.1 | 0.870 | 66.8 | 6.73 | 86.4 | *We applied the last four Standard sft, SteerLM, Rejection Sampling and DPO, and found DPO to be most performant ## Ethical Considerations This model was developed with a focus on mitigating modality conflict and catastrophic forgetting in MLLMs. Users are encouraged to consider the potential biases and limitations inherent in the training data and model outputs, especially when deploying the model in diverse and sensitive contexts. ## Limitations - The model's training dataset, while addressing key gaps in VQA datasets, is relatively small at 6k entries. This may limit the model's generalizability across broader or more diverse multi-modal tasks. - Performance enhancements, particularly in language instruction capabilities post-visual tuning, are based on the current scope of evaluated benchmarks and datasets. The model's efficacy may vary in different or more challenging contexts. ## Acknowledgments This work was made possible through the contributions of Shengzhi Li, Rongyu Lin, and Shichao Pei, and supported by their respective institutions. ## Citation Please cite this work as: ```bibtex @misc{li2024multimodal, title={Multi-modal preference alignment remedies regression of visual instruction tuning on language model}, author={Shengzhi Li and Rongyu Lin and Shichao Pei}, year={2024}, eprint={2402.10884}, archivePrefix={arXiv}, primaryClass={cs.CL} }
许可证:Apache-2.0 [[Arxiv论文](https://arxiv.org/abs/2402.10884)] [[GitHub仓库](https://github.com/findalexli/mllm-dpo)] [[数据集](https://huggingface.co/datasets/alexshengzhili/llava-preference-dpo)] [[模型](https://huggingface.co/alexshengzhili/llava-v1.5-13b-dpo/edit/main/README.md)] [[数据集](https://huggingface.co/datasets/alexshengzhili/mllm-dpo)] **论文发布地址:** [https://arxiv.org/abs/2402.10884](https://arxiv.org/abs/2402.10884) **开发者:** 李盛智(TIFIN)、林容宇(阿卜杜拉国王科技大学,KAUST)、裴世超(马萨诸塞大学波士顿分校) **所属机构:** TIFIN、阿卜杜拉国王科技大学(KAUST)、马萨诸塞大学波士顿分校 **联系方式:** alex.li@tifin.com, rongyu.lin@kaust.edu.sa, shichao.pei@umb.edu ## 概述 llava-v1.5-13b-dpo模型及对应数据集旨在增强多模态大语言模型(Multi-modal Large Language Model, MLLM)的指令遵循能力,尤其针对视觉指令微调可能降低语言能力的场景。本模型采用了一种新颖的直接偏好优化(Direct Preference Optimization, DPO)方法,并结合了精心构建的6000条规模的视觉问答(Visual Question Answering, VQA)偏好数据集,从而在多模态任务与基准测试中取得更优性能。 ## 预期用途 - **核心应用场景**:本模型适用于需要融合文本与图像模态的任务,包括但不限于视觉问答(VQA)、图像字幕生成以及多模态指令遵循。 - **目标受众**:自然语言处理、计算机视觉及多模态人工智能领域的研究人员与从业者。 ## 训练数据 MM-LLM-DPO模型使用一个轻量化的(6000条数据)VQA偏好数据集进行训练,该数据集针对答案从5个质量维度进行了精细化标注。本数据集旨在弥补现有VQA数据集普遍存在的多样性与复杂度缺口。 ## 模型评估 本模型在多项基准测试中显著优于Vicuna与LLaVA等基线模型: - **MT-Bench**:得分达6.73,超过Vicuna的6.57与LLaVA的5.99。 - **视觉指令性能**:在MM-Vet基准上提升4.9%,在LLaVA-Bench基准上提升6%。 | 模型名称 | MM-Vet | LLaVA-bench | PoPe | MM-Bench | MT-bench | AlpacaEval | |------------|--------|-------------|-------|----------|----------|------------| | **Vicuna-1.5-13b [16]** | - | - | - | - | 6.57 | 81.4 | | **LLaVA-1.5-13b [10]** | 36.3 | 73.1 | 0.859 | 67.4 | 5.99 | 79.3 | | **LLaVA-RLHF-13b [23]**| 37.2 | 76.8 | 0.869 | 60.1 | 6.18 | 81.0 | | **标准监督微调(Standard SFT)** | 36.5 | 63.7 | 0.850 | 65.4 | 5.01 | 50.2 | | **SteerLM** | 35.2 | 67.0 | 0.878 | 65.1 | 5.70 | 68.8 | | **拒绝采样(Rejection-sampling)** | 38.0 | 70.6 | 0.883 | 67.6 | 6.22 | 74.9 | | **llava-v1.5-13b-dpo** | 41.2 | 79.1 | 0.870 | 66.8 | 6.73 | 86.4 | *我们对标准监督微调(Standard SFT)、SteerLM、拒绝采样(Rejection Sampling)与DPO四种方法进行了对比实验,结果显示DPO的综合性能最优。 ## 伦理考量 本模型的开发聚焦于缓解多模态大语言模型中的模态冲突与灾难性遗忘问题。我们鼓励用户在将模型部署于多样化且敏感的场景时,充分考虑训练数据与模型输出中潜在的偏见与局限性。 ## 局限性 - 本模型的训练数据集虽弥补了VQA数据集的部分关键缺口,但规模仅为6000条,相对较小,这可能限制模型在更广泛或更多样化的多模态任务中的泛化能力。 - 性能提升(尤其是视觉微调后语言指令能力的提升)基于当前评估的基准测试与数据集范围,模型在不同或更具挑战性的场景中的效果可能存在差异。 ## 致谢 本研究得益于李盛智、林容宇与裴世超的贡献,并得到其所属机构的支持。 ## 引用格式 请按以下格式引用本工作: bibtex @misc{li2024multimodal, title={Multi-modal preference alignment remedies regression of visual instruction tuning on language model}, author={Shengzhi Li and Rongyu Lin and Shichao Pei}, year={2024}, eprint={2402.10884}, archivePrefix={arXiv}, primaryClass={cs.CL} }
数据集概述
模型与数据集设计
- 目的: 提升多模态大型语言模型(MLLMs)的指令遵循能力,特别是在视觉指令调优可能降低语言熟练度的场景中。
- 技术: 采用直接偏好优化(DPO)方法和6K条目VQA偏好数据集。
预期用途
- 主要应用: 文本和图像模态整合任务,如视觉问答(VQA)、图像标注和多模态指令遵循。
- 目标用户: 自然语言处理、计算机视觉和多模态AI领域的研究者和实践者。
训练数据
- 数据集: 使用6K条目的轻量级VQA偏好数据集,针对5个质量指标进行细致标注。
评估
- 性能提升: 在多个基准测试中超越基线模型,如Vicuna和LLaVA。
- 具体成绩: 在MT-Bench上得分6.73,MM-Vet上提升4.9%,LLaVA-Bench上提升6%。
伦理考量
- 关注点: 用户应考虑训练数据和模型输出中可能存在的偏见和局限性,特别是在多样化和敏感环境中部署时。
限制
- 数据集大小: 6K条目的数据集可能限制模型在更广泛或多样的多模态任务中的泛化能力。
- 性能评估: 性能提升基于当前评估的基准和数据集,可能在不同或更具挑战性的环境中表现不同。
致谢
- 贡献者: Shengzhi Li, Rongyu Lin, Shichao Pei。
- 支持机构: TIFIN, KAUST, University of Massachusetts Boston。
引用信息
- 引用格式: bibtex @misc{li2024multimodal, title={Multi-modal preference alignment remedies regression of visual instruction tuning on language model}, author={Shengzhi Li and Rongyu Lin and Shichao Pei}, year={2024}, eprint={2402.10884}, archivePrefix={arXiv}, primaryClass={cs.CL} }




