icebearai/llm-and-more

LLM-And-More is a professional, plug-and-play, llm trainer and application builder that guides you through the complete LLM workflow from data to evaluation, from training to deployment, from idea to sevice. / LLM-And-More 是一个专业、开箱即用的大模型训练及应用构建一站式解决方案，包含从数据到评估、从训练到部署、从想法到服务的全流程最佳实践。

llm-and-more's Issues

什么时候支持一下llama3的微调和评估？

如题

你好，微信群二维码过期了

如题

增加詳細的英文文檔支持

問題描述

首先，非常感謝項目團隊開發了這麼一個功能強大且易於使用的LLM應用構建方案。目前，我注意到項目的README及大部分文檔都是以中文提供的，雖然這對於中文用戶非常友好，但對於非中文使用者來說可能會構成一定的障礙。鑑於LLM的廣泛應用和國際化趨勢，我認爲增加英文文檔是非常必要的。

建議解決方案

1. 增加英文README: 項目可以首先提供一份英文版的README文件，涵蓋項目概述、安裝指南、使用示例等基礎但重要的信息。
2. 逐步完善英文文檔: 對於項目中的核心功能，如數據模塊、訓練模塊等，逐步提供詳細的英文文檔，幫助更多的國際開發者和用戶理解並使用這個項目。
3. 社區參與: 鼓勵社區成員參與文檔翻譯和校對工作，既能加速文檔的國際化進程，又能增強社區的參與感和凝聚力。可以通過GitHub Issue或Pull Request的方式來組織這項工作。

預期效果

• 擴大用戶羣體: 英文文檔將幫助項目吸引更多的國際用戶，尤其是非中文環境的開發者和企業。
• 增強項目影響力: 國際化的文檔不僅能提升項目的可訪問性，還有助於提高項目的國際知名度和影響力。
• 促進社區多樣性: 鼓勵更多的國際用戶參與到項目貢獻中來，不僅可以提高文檔質量，還能促進社區成員背景的多樣性。

結論

LLM-And-More是一個有巨大潛力的項目，我相信通過增強文檔的國際化，能夠讓這個項目更加強大，惠及全球更多的用戶和開發者。期待您的回覆和看法！

添加支持多任务训练的功能

LLM-And-More 目前在训练模块中只支持针对单一任务进行模型训练,这在某些情况下可能无法满足实际需求。

很多场景下,我们希望训练出一个能够同时处理多个任务的通用模型,比如在FAQ机器人中既能回答常见问题,又能进行简单的文本生成和情感分析。因此,，建议在训练模块中增加支持多任务训练的功能,具体建议如下:

1. 支持用户自定义多个训练任务,并设定每个任务的损失函数和评估指标。
2. 在训练过程中,能够显示各个任务的实时loss曲线和评估指标,帮助用户监控训练进度。
3. 提供智能的任务权重调整功能,帮助用户平衡不同任务的重要性,提高模型在各个任务上的综合性能。
4. 在部署模块中,支持用户选择在线部署单一任务或者多任务的模型。

这些功能的增加,将大大提高 LLM-And-More 的适用性,使其能够满足更广泛的业务需求。希望开发团队能够考虑并实现这些改进建议,谢谢。

chat模型五维图评估问题

如上图所示，chat 类的模型评估只有推理能力一条。其他能力无法评估

Improve Model Training Explainability and Debugging Features

While training models with LLM-And-More, I've observed that enhancing the explainability and debugging capabilities during the training process could significantly improve user experience and development efficiency. Here are my specific suggestions:

1. Implement Explainability Tools for Training: Understanding the learning process and decision-making basis of the model is crucial for developers. It's recommended to add visualization tools to LLM-And-More that display key metrics, changes in loss functions, gradient information, etc., during training. This would help developers intuitively comprehend the model's training status and performance.

2. Support Debugging Features: Developing and debugging models often require experimenting with different hyperparameters, data preprocessing methods, etc., to optimize model performance. It's suggested to introduce debugging features in LLM-And-More, such as an interactive debugging interface that allows users to observe the model's predictions on small sample data in real-time, facilitating prompt identification and resolution of issues.

3. Provide Model Explainability Tools: For large-scale models, understanding the decision-making process and the explainability of predictions is crucial. It's recommended to incorporate explainability tools into LLM-And-More, such as visualizing feature importance, generating explanatory texts, etc., to help users understand the principles and logic behind model predictions.

These improvements would aid in enhancing users' understanding and control over the model training process, accelerating the debugging and optimization stages. I hope the LLM-And-More team will consider and implement these features to further increase the project's practicality and user-friendliness. Thank you for your attention to developers' needs!

关于增强模型适配性与部署灵活性的建议

首先感谢团队开发的LLM-And-More项目，该项目为大模型训练及应用构建提供了一站式解决方案，极大简化了从数据处理到模型评估、从训练到部署的整个流程。

经过一段时间的使用，我发现项目在模型适配性和部署灵活性方面还有提升的空间。以下是我个人的一些建议，希望能够对项目的进一步发展有所帮助——

目前LLM-And-More项目已经支持了多种模型，如Baichuan2、ChatGLM3、LLaMA等，这为用户提供了广泛的选择。然而，考虑到大模型领域的快速发展，新的模型和优化算法不断出现，建议项目能够：

1. 增加更多模型的支持：定期评估和集成新出现的高效或特色模型，以保持项目的领先性和吸引力。
2. 提供模型适配指南：为用户提供详细的模型适配指南，包括如何添加新模型、调整模型参数等，让有能力的用户能够自行扩展模型库。

项目已经提供了Docker-compose的部署方式，这对于大多数用户来说已经足够方便。但对于需要在特定环境（如没有Docker环境的服务器、边缘计算设备等）部署模型的用户来说，可能会遇到困难。因此，我建议：

1. 增加无Docker部署方案：提供基于虚拟环境（如Python虚拟环境）的部署方案，使项目能够在更广泛的环境中运行。
2. 提供详细的部署文档：针对不同的部署环境（云服务器、本地服务器、边缘设备等），提供更详细的部署指南和最佳实践，帮助用户解决部署过程中可能遇到的问题。

希望这些建议能够对LLM-And-More项目的未来发展有所帮助，期待项目能够不断进步，为更多用户提供便利。

再次感谢团队的努力和贡献！

[Feature Request] Custom Model Training

I've noted that LLM-And-More stands out as a powerful one-stop solution for Large Language Models (LLMs), offering a comprehensive workflow from data handling to evaluation, and from training to deployment. However, I believe there's room for enhancement by introducing custom model training capabilities, which would afford greater flexibility and customization options.

Currently, LLM-And-More appears to prioritize providing pre-defined, high-performance models along with default parameters for an out-of-the-box training experience. Nonetheless, certain users may seek to train their own models to meet specific needs or research objectives.

Hence, I propose integrating a feature within LLM-And-More that allows users to upload and utilize their own pre-trained models or custom model architectures. Such functionality would offer users the following advantages:

1. Flexibility: Users could select the model architecture and parameter settings that best suit their needs, leading to improved performance and adaptability.
2. Customization: Researchers and developers, who may have trained specific models in other projects, could directly use these models within LLM-And-More for further training and deployment.
3. Innovation: Allowing users to upload custom models could transform LLM-And-More into a more open and innovative platform, encouraging the sharing and exploration of various model architectures and techniques.

I believe that introducing the custom model training capability will further enhance the value and appeal of LLM-And-More, catering to a broader range of user requirements. I hope the team will consider this suggestion and look into implementing this feature in future releases.

Thank you for your dedication and contributions to the open-source community!