# Model Fine-tuning

FastApply-7B-v1.0 is a large language model specifically designed for code editing tasks. Built on the Qwen2.5 Coder architecture, it's fine-tuned for fast and accurate code modifications. This model rapidly generates complete file edits, supporting immediate code application tasks, making it ideal for integration into AI-powered code editors. It demonstrates high throughput and high editing accuracy in deployment, achieving speeds of approximately 150 tokens/second. Developed by Kortix and licensed under Apache-2.0, it aims to support data generation and model fine-tuning through a fast application process.

Bakery is an online platform focused on the fine-tuning and monetization of open-source AI models, providing a handy tool for AI startups, machine learning engineers, and researchers to easily fine-tune AI models and monetize them in the market. Its key advantages lie in a user-friendly interface and powerful features, allowing users to quickly create or upload datasets, adjust model settings, and monetize in the marketplace. Bakery’s background indicates that it aims to advance the development of open-source AI technology while offering developers more business opportunities. Although specific pricing information isn’t clearly displayed on the page, its position is to provide an efficient tool for professionals in the AI field.

The Zhipu Qingliu AI Open Platform is an enterprise-level AI agent development platform that utilizes Zhipu large model technology to help enterprises quickly build professional-grade agents for rapid application in business scenarios. The platform provides functionalities such as model services, agent construction, data security, effectiveness evaluation, and system integration, supporting enterprises in protecting data through intranet deployment and local storage to ensure data security and intellectual property rights. The Zhipu AI Open Platform, with its leading technology, flexible workflow orchestration, autonomous invocation of enterprise-defined data and tools, and mature AI-native application experience, serves as a significant enabler of enterprise digital transformation.

SFR-Judge is a series of evaluation models launched by Salesforce AI Research, aimed at accelerating the evaluation and fine-tuning processes of large language models (LLMs) through artificial intelligence technology. These models can perform a variety of evaluation tasks, including pairwise comparisons, single-item scoring, and binary classification, while providing explanations to avoid black-box issues. SFR-Judge has demonstrated exceptional performance in multiple benchmark tests, proving its effectiveness in evaluating model outputs and guiding fine-tuning.

diffusion-e2e-ft is an open-source fine-tuning tool for image-conditioned diffusion models that enhances performance on specific tasks by fine-tuning pre-trained diffusion models. This tool supports various models and tasks, such as depth estimation and normal estimation, and provides detailed usage instructions and model checkpoints. It has significant applications in image processing and computer vision, enabling substantial improvements in model accuracy and efficiency for specific tasks.

Prem is an AI platform that accelerates the development of personalized LLMs (Small Language Models). It offers a user-friendly development platform for building generative AI solutions. Prem also provides customizable model fine-tuning and training services to meet unique business needs. Through Prem, users can gain maximum control and ownership over AI's potential without requiring deep technical expertise.

The Llama Chinese Community is a technical community focused on optimizing the Llama model in Chinese and upper-level construction. The community provides pre-trained models based on large-scale Chinese data and continuously iterates and upgrades the Chinese capabilities of the Llama2 and Llama3 models. With a team of senior engineers, rich community activities, and an open and shared cooperative environment, the community aims to promote the development of Chinese natural language processing technology.

XTuner is a high-performance, flexible, and feature-rich fine-tuning toolkit designed for large models such as InternLM, Llama, Baichuan, Qwen, and ChatGLM. It supports the pre-training and fine-tuning of LLM and VLM on almost all GPUs, and can automatically schedule high-performance operations like FlashAttention and Triton kernels to enhance training throughput. XTuner is compatible with DeepSpeed and supports various ZeRO optimization techniques. It also supports various LLMs and VLMs, such as LLaVA, and has been designed with robust data pipelines to accommodate any format of datasets. Moreover, XTuner supports multiple training algorithms, including QLoRA, LoRA, and full-parameter fine-tuning, allowing users to choose the solution that best fits their needs.

The MindEye2 project demonstrates how, by pre-training a model on multiple subjects using only 1 hour of fMRI data, fine-tuning can be performed on new subjects to achieve visual perception reconstruction.

Cappy is a novel approach designed to improve the performance and efficiency of large, multi-task language models. It is a lightweight, pre-trained scoring model based on RoBERTa, with only 360 million parameters. Cappy can independently solve classification tasks or act as an auxiliary component to enhance the performance of language models. Fine-tuning Cappy on downstream tasks effectively integrates supervisory information, improving model performance. This process does not require backpropagation to the language model parameters, reducing memory requirements. Cappy is applicable to both open-source and closed-source language models, providing an efficient model fine-tuning method.

SCEdit is an efficient generation model fine-tuning framework proposed by Alibaba. It enhances the fine-tuning capability for downstream text-to-image generation tasks and enables fast adaptation to specific generation scenarios. Compared to LoRA, it can save 30%-50% of training memory costs. Moreover, it can be directly extended to controllable image generation tasks, requiring only 7.9% of the parameter amount of ControlNet conditional generation and saving 30% of memory usage. It supports various conditional generation tasks, including edge maps, depth maps, segmentation maps, poses, color maps, and image completion.

LLaVA-3b is a model fine-tuned based on Dolphin 2.6 Phi, using the SigLIP 400M visual tower in an LLaVA manner. The model features multiple image labels and outputs from the latest layer of the visual encoder. This model is based on Phi-2 and is subject to the Microsoft Research license, prohibiting commercial use. Thanks to ML Collective for providing computational resource credits.

Windows AI Studio streamlines the development of generative AI applications by bringing together cutting-edge AI development tools and models from Azure AI Studio and other repositories like Hugging Face. It empowers developers to fine-tune, customize, and deploy the latest small language models (SLMs) within their Windows applications. It offers end-to-end guided workspace setup, including a model configuration UI and guided steps, to fine-tune popular SLMs like Phi and state-of-the-art models such as Llama 2 and Mistral.