The paper "Deep Learning Model Reuse in the HuggingFace Community: Challenges, Benefit and Trends" by Mina Taraghi, Gianolli Dorcelus, Armstrong Foundjem, Florian Tambon, and Foutse Khomh explores the challenges and benefits associated with the reuse of Pre-Trained Models (PTMs) within the HuggingFace community. The study employs a mixed-methods approach, combining qualitative and quantitative analysis to provide a comprehensive understanding of PTM reuse. - **Challenges:** - Limited guidance for beginner users. - Struggles with model output comprehensibility in training or inference. - Lack of model understanding. - Training pipeline issues. - Memory and performance problems. - Challenges with using platforms and libraries other than HF. - Documentation and tutorial deficiencies. - Specific solution needs. - Dataset acquisition and usage issues. - Incomprehensible outputs. - Requests and questions related to HF libraries' features. - **Benefits:** - Expert-provided solutions and clarifications. - Collaborations among users and experts. - Acknowledgments of help. - Call for collaboration topics. - Announcements and cross-disciplinary contributions. - **BERT** is the most discussed and uploaded model, maintaining its position over time. - **T5**, **RoBERTa**, and **GPT2** are also popular. - There is a negative correlation between the trend of some models provided on the hub and the trend of discussion about these models. - Despite the introduction of tools for documenting models, the quantity of model documentation has not improved over time. - **Beginner Users:** Need improved tutorial examples and model guidance. - **Model Documentation:** Improve the quantity and quality of model cards to enhance model understanding and selection. - **API Usage and Tools:** Provide more detailed documentation and simplify the interface for searching within model documentation. - **Model Providers:** Consider trends in model usage and provide models that align with community needs. - **Pan et al. [14]:** Studied GitHub repositories of popular models supported by HF transformers. - **McMillan-Major et al. [72]:** Proposed documentation for language datasets and NLG models based on HuggingFace Hub and GEM benchmark. - **Jiang et al. [1]:** Assessed artifacts and security features across 8 model hubs. - **Jiang et al. [12]:** Conducted interviews with HF practitioners and analyzed PTM packages on HF. - **Shen et al. [69]:** Introduced a tool, *HuggingGPT*, to manage multiple HF models and find solutions for specific tasks. - **You et al. [17]:** Proposed a method for ranking PTMs based on transferability metrics and fine-tuning using Bayesian procedures. - **Costano et al. [73]:** Analyzed carbon-footprint reportsThe paper "Deep Learning Model Reuse in the HuggingFace Community: Challenges, Benefit and Trends" by Mina Taraghi, Gianolli Dorcelus, Armstrong Foundjem, Florian Tambon, and Foutse Khomh explores the challenges and benefits associated with the reuse of Pre-Trained Models (PTMs) within the HuggingFace community. The study employs a mixed-methods approach, combining qualitative and quantitative analysis to provide a comprehensive understanding of PTM reuse. - **Challenges:** - Limited guidance for beginner users. - Struggles with model output comprehensibility in training or inference. - Lack of model understanding. - Training pipeline issues. - Memory and performance problems. - Challenges with using platforms and libraries other than HF. - Documentation and tutorial deficiencies. - Specific solution needs. - Dataset acquisition and usage issues. - Incomprehensible outputs. - Requests and questions related to HF libraries' features. - **Benefits:** - Expert-provided solutions and clarifications. - Collaborations among users and experts. - Acknowledgments of help. - Call for collaboration topics. - Announcements and cross-disciplinary contributions. - **BERT** is the most discussed and uploaded model, maintaining its position over time. - **T5**, **RoBERTa**, and **GPT2** are also popular. - There is a negative correlation between the trend of some models provided on the hub and the trend of discussion about these models. - Despite the introduction of tools for documenting models, the quantity of model documentation has not improved over time. - **Beginner Users:** Need improved tutorial examples and model guidance. - **Model Documentation:** Improve the quantity and quality of model cards to enhance model understanding and selection. - **API Usage and Tools:** Provide more detailed documentation and simplify the interface for searching within model documentation. - **Model Providers:** Consider trends in model usage and provide models that align with community needs. - **Pan et al. [14]:** Studied GitHub repositories of popular models supported by HF transformers. - **McMillan-Major et al. [72]:** Proposed documentation for language datasets and NLG models based on HuggingFace Hub and GEM benchmark. - **Jiang et al. [1]:** Assessed artifacts and security features across 8 model hubs. - **Jiang et al. [12]:** Conducted interviews with HF practitioners and analyzed PTM packages on HF. - **Shen et al. [69]:** Introduced a tool, *HuggingGPT*, to manage multiple HF models and find solutions for specific tasks. - **You et al. [17]:** Proposed a method for ranking PTMs based on transferability metrics and fine-tuning using Bayesian procedures. - **Costano et al. [73]:** Analyzed carbon-footprint reports