AI Trending

Project directory

Projects

Open-source AI projects tracked by the leaderboard, ordered by GitHub stars.

Python 5 stars

NvMayMay/nvfp4-lora-spark

LoRA fine-tune and serve NVFP4 models on one DGX Spark (GB10, 128 GB UMA): text backbones via generic-family onboarding, plus VLMs (vision tower, or LLM+tower jointly via --train-target both) validated end-to-end on Pixtral and Nemotron-Omni. Fused Triton dequant; runtime-LoRA and merge serving.

### Overview: nybbloris is a project that fine-tunes and serves NVFP4 models on a single DGX Spark, leveraging runtime LoRA and a generic fallback mechanism to ensure the adapter remains functional and validated across various target models and architectures. ### Key Features: - Fine-tune a 100B+ MoE on the 4-bit weights NVIDIA ships. - Train LoRA directly on the served 4-bit weights without the need for a BF16 base. - Serve runtime-LoRA on NVFP4, FP8, and bf16 models. - Proven end-to-end on Command-A, achieving a -28.8 -> -17.4, +11.4 nats improvement in base vs adapter. - Fine-tune the vision tower of a VLM and train the LLM backbone and tower jointly. - Use a generic fallback mechanism to validate and serve models from unregistered checkpoints. ### Tech Stack: Python ### Why it Matters: nybbloris provides a reliable and validated solution for fine-tuning and serving NVFP4 models on a single DGX Spark, ensuring the adapter remains functional and the model's performance is maintained across various configurations and architectures.
Python 5 stars

OpenMOSS/sglang

An OpenMOSS fork of SGLang with serving support for the MOSS-TTS family and MOSS-Audio

Overview: This repository is an OpenMOSS fork of SGLang, specifically designed to support the MOSS-TTS Family and MOSS-Audio, including models such as MOSS-Audio, MOSS-TTS (Delay), MOSS-SoundEffect, MOSS-TTSD v1.0, and MOSS-TTSD v0.7. Key features: - MOSS-Audio - MOSS-TTS (Delay) / MOSS-SoundEffect - MOSS-TTSD v1.0 - MOSS-TTSD v0.7 Tech stack: Python Why it matters: This repository provides comprehensive support for the MOSS-TTS Family and MOSS-Audio models, enhancing the functionality of SGLang for audio understanding and generation tasks. Repository: OpenMOSS/sglang GitHub stars: 5 GitHub description: An OpenMOSS fork of SGLang with serving support for the MOSS-TTS Family and MOSS-Audio.
Jupyter Notebook 5 stars

Pavansomisetty21/A-Fine-Tuned-Model-for-Medical-Named-Entity-Recognition-using-Gemini-LLM

In this we finetuned the Gemini model with our own medical NER dataset and used to recognize Name Entities

### Overview: This project fine-tunes the Gemini model with a medical named entity recognition dataset to accurately identify and classify medical entities in text. ### Key Features: - Fine-tuning of the Gemini model for medical named entity recognition. - Integration of a medical dataset for improved accuracy. - Extraction of medical entities such as diseases, symptoms, drugs, and body parts. - Output format: Entities with their corresponding labels. ### Tech Stack: Jupyter Notebook ### Why it Matters: This project is crucial for healthcare professionals and researchers who need to extract structured information from unstructured medical text, enhancing their ability to analyze and understand patient records effectively.
Jupyter Notebook 5 stars

Pavansomisetty21/Visual-Question-Answering-Pixtral_Vision_Finetuning_Unsloth

In this we finetune Pixtral-12B-2409 model using unsloth for visual Question Answering(NLP Task)

Overview: This project involves finetuning the Pixtral-12B-2409 model using the unsloth package for visual Question Answering (NLP Task). Key features: Finetuning on the unsloth/llava-instruct-mix-vsft-mini dataset with the Pixtral-12B-2409 model. Tech stack: Jupyter Notebook Why it matters: This project aims to enhance the performance of visual Question Answering by fine-tuning a pre-trained model on a specific dataset, potentially improving accuracy and relevance in NLP tasks.
Jupyter Notebook 5 stars

PxA-Labs/safetypulse

Early warning for LLM safety drift during fine-tuning: a Circuit Integrity Score read from the model's refusal circuit flags unsafe checkpoints before harm is measurable, with SafeTune repair and AuditKIT audit.

### Overview: SafetyPulse is a tool designed to detect safety drift during the fine-tuning of large language models (LLMs). It uses a Circuit Integrity Score (CIS) to monitor the model's refusal mechanism, which is cheaper and faster than a full red-team audit. ### Key Features: - **Circuit Integrity Score (CIS):** A score computed from 50 prompt pairs to gauge the model's refusal mechanism integrity. - **SafeTune Repair:** A repair mechanism that can be used to correct unsafe checkpoints identified by SafetyPulse. - **AuditKIT Audit:** An audit tool to verify the model's safety status. ### Tech Stack: Python 3.10+ with Jupyter Notebook ### Why It Matters: Early detection of safety drift is crucial for maintaining the integrity and safety of LLMs, preventing harmful outputs before they become measurable. SafetyPulse provides a cost-effective solution to ensure model safety during fine-tuning.
Python 5 stars

R0U5/Reforge

A modular LoRA fine-tuning pipeline for Hugging Face models with built-in data cleaning, smart early stopping, automatic mode detection, and cross-platform support.

### Overview: Reforge is a modular LoRA fine-tuning pipeline for Hugging Face models that offers built-in data cleaning, smart early stopping, automatic mode detection, and cross-platform support. It includes a Tkinter dashboard for interactive control and is designed to handle various datasets, including multimodal image+text training. ### Key features: - **Automatic Training Mode Detection** — selects the appropriate pipeline based on dataset columns. - **Built-in Data Cleaning** — cleans LaTeX, HTML, markdown, and other artifacts from text data. - **Model Profiles** — includes Phi-2, Phi-3, Phi-3-Vision, and more for pre-tuned configurations. - **Smart Early Stopping** — detects loss plateaus and uses exposure floors, LR gates, and dynamic hard caps for efficient training. - **Automatic Sequence Length** — measures token length to optimize model performance. ### Tech Stack: Python ### Why it Matters: Reforge simplifies the process of fine-tuning Hugging Face models, ensuring high-quality data cleaning, smart training, and efficient early stopping, making it an essential tool for developers working with various datasets and models.
Python 5 stars

R4255/converse

AI command layer for a grocery app: a local LLM agent that drives the UI, with hybrid search, outcome-based evals (74% → 93%), and a LoRA fine-tuned 1.5B model

### Overview: Converse is an AI command layer for a grocery app that integrates with local models, Postgres, and AWS services. It allows users to perform various actions directly from the app's interface, ensuring all interactions are reversible and fine-tuned with a 1.5B model. ### Key features: - **Hybrid Search**: Combines local and remote model responses for a seamless search experience. - **Outcome-Based Evaluations**: Uses a 130-task main suite and a 40-task held-out set to evaluate agent performance. - **Prompt Grounding**: Ensures that the agent asks for user confirmation before executing any irreversible actions. - **Fine-Tuned Model**: A 1.5B model fine-tuned with LoRA to match a 3B model, improving accuracy and efficiency. ### Tech stack: Python ### Why it matters: Converse is essential for a grocery app because it ensures a smooth and user-friendly experience, improving the overall user satisfaction and engagement. The fine-tuned model and prompt grounding mechanisms make the app more reliable and efficient, leading to better performance and user trust. ### Repository: R4255/converse ### GitHub stars: 5 ### GitHub description: AI command layer for a grocery app: a local LLM agent that drives the UI, with hybrid search, outcome-based evals (74% → 93%), and a LoRA fine-tuned 1.5B model.
TypeScript 5 stars

Samashi47/FSTT-LLM-Chat-Bot

A chat bot for the Faculty of Sciences and Technologies of Tangier. Using both RAG and finetuning as separate models.

Overview: A chat bot for the Faculty of Sciences and Technologies of Tangier, utilizing both RAG and fine-tuning models. Key features: - Fine-tuned Llama3-8B for public information from the faculty's website. - RAG implementation using Solon-embeddings-large-0.1 for embeddings and Llama3-8B for inference. Tech stack: TypeScript Why it matters: This project aims to enhance educational interactions by providing a more relevant and contextually aware chatbot, improving student and faculty experiences in an academic setting.
Jupyter Notebook 5 stars

SatvikPraveen/FashionMNIST-Analysis

Reproducible empirical study of Fashion-MNIST classification in PyTorch: what modern training recipes, pretrained backbones (ViT-Tiny, ConvNeXt) and seed ensembles buy, and what they cost in calibration and robustness. Every number is a mean over 3 to 5 seeds with 95% CIs and paired tests, with per-run provenance.

Overview: This project is a reproducible empirical study of image classification on the Fashion-MNIST dataset using PyTorch. The study evaluates various models and training recipes, including pretrained backbones and seed ensembles, to understand their impact on model accuracy, calibration, and robustness. Key features: - Pretrained ViT-Tiny/16 (224 px) - Pretrained ConvNeXt-Tiny - 5-seed ensemble of TinyVGG, no random crop - ResNet-18 (custom, from scratch) - TinyVGG, no random crop - MiniCNN, no random crop - TinyVGG, default recipe - MiniCNN, default recipe - Best classical model (kNN on PCA features) Tech stack: Jupyter Notebook Why it matters: This study provides insights into the performance of various image classification models and training techniques on a small grayscale benchmark, highlighting the importance of pretraining, seed ensembles, and augmentation strategies in achieving high accuracy, calibration, and robustness. Repository: SatvikPraveen/FashionMNIST-Analysis Primary language: Jupyter Notebook GitHub stars: 5 GitHub description: Reproducible empirical study of Fashion-MNIST classification in PyTorch: what modern training recipes, pretrained backbones (ViT-Tiny, ConvNeXt) and seed ensembles buy, and what they cost in calibration and robustness. Every number is a mean over 3 to 5 seeds with 95% CIs and paired tests, with per
Python 5 stars

Shreyjain203/Wikipedia-Continual-Learning-RAG

This repository implements a self-updating RAG (Retrieval Augmented Generation) model. It leverages Wikipedia for factual grounding and can fine-tune itself when information is unavailable. This allows the model to continually learn and adapt, offering a dynamic and informative response.

Overview: This project implements a self-updating RAG (Retrieval Augmented Generation) model that leverages Wikipedia for factual grounding. It fine-tunes the model when information is unavailable, allowing for dynamic and informative responses. Key features: - Continual learning - Factual grounding from Wikipedia - Fine-tuning when information is missing Tech stack: Python Why it matters: This project offers a dynamic and informative response by continuously learning from Wikipedia, making it useful for applications requiring accurate and up-to-date information. Repository: Shreyjain203/Wikipedia-Continual-Learning-RAG Primary language: Python GitHub stars: 5 GitHub description: This repository implements a self-updating RAG (Retrieval Augmented Generation) model that leverages Wikipedia for factual grounding and can fine-tune itself when information is unavailable. This allows the model to continually learn and adapt, offering a dynamic and informative response.
Python 5 stars

SpectreSpect/ConversationalGPT-2

ConversationalGPT-2 is a transformer-based language model built for engaging in natural dialogues. It can handle casual conversations, recall facts, and understand basic world knowledge. Trained in PyTorch, the model is available for testing and use.

### Overview: ConversationalGPT-2 is a transformer-based language model designed for engaging in natural dialogues. It can handle casual conversations, recall facts, and understand basic world knowledge. ### Key features: - **Natural Conversations**: Engaging in casual and everyday conversations. - **Facts Recall**: Remembering and recalling factual information. - **World Knowledge**: Understanding basic world knowledge and common sense. ### Tech stack: - **Python**: The primary programming language used for development. - **PyTorch**: The deep learning framework used for training the model. - **PyTorchXLA**: Accelerated training using XLA (Accelerated Linear Algebra). - **CometML**: For model deployment and integration. - **TPU**: Training on the TPU (Tensor Processing Unit). ### Why it matters: ConversationalGPT-2 is crucial for applications requiring natural language interaction, such as chatbots, customer service, and educational tools. Its ability to handle casual conversations, recall facts, and understand basic world knowledge makes it highly valuable in various domains.
Python 5 stars

SpencerSzabados/Fine-tune-Stable-Diffusion-VAE

A quick reference for fine-tuning Stable Diffusions VAE on custom datasets.

Overview: This repository provides detailed instructions for fine-tuning Stable Diffusion's Variational Auto-Encoder on custom image datasets. It includes scripts for training at a resolution of 512x512x3 and supports modifications for custom resolutions. Key features: - Fine-tuning Stable Diffusion's Variational Auto-Encoder on custom datasets - Scripts for training at 512x512x3 resolution - Options for custom resolution training - Environment setup using conda and requirements.yaml - Launch options for 'fine-tune.sh' script - Default working parameters for training Tech stack: Python Why it matters: This repository offers a practical guide for fine-tuning Stable Diffusion's VAE, making it useful for developers looking to experiment with custom data or fine-tune models for specific applications. Repository: SpencerSzabados/Fine-tune-Stable-Diffusion-VAE GitHub stars: 5 GitHub description: A quick reference for fine-tuning Stable Diffusions VAE on custom datasets.
Jupyter Notebook 5 stars

StepanTita/llm-friend-bot

This project seeks to fine-tune the GPT-2.5 model with a personal touch. By training it on personal Telegram chats, we aim to capture the essence of individualized writing styles. To spice things up, we've also fed the model a sprinkle of anecdotes and a dash of math tasks.

Overview: This project aims to fine-tune the GPT-2.5 model with personal Telegram chat data to create a model that captures individual writing styles, incorporates humor, and solves math tasks. Key features: - Captures individual writing styles - Adds a sense of humor - Solves math tasks Tech stack: Python, PyTorch, Hugging Face Transformers Why it matters: This project offers a unique model that can generate personalized writing, add humor, and handle mathematical tasks, making it useful for various applications.
Python 5 stars

TammineniTanay/distributed-finetune-pipeline

End-to-end distributed LLM fine-tuning pipeline: data curation → QLoRA + DPO training → TIES/DARE/SLERP model merging → evaluation → vLLM deployment. Built with DeepSpeed ZeRO-3, Flash Attention, Prometheus monitoring, and AWS spot instance infrastructure.

### Overview: This project is an end-to-end distributed LLM fine-tuning pipeline that includes data curation, QLoRA + DPO training, model merging, evaluation, and deployment. Built with DeepSpeed ZeRO-3, Flash Attention, Prometheus monitoring, and AWS spot instance infrastructure, it is adaptable for any HuggingFace-compatible model. ### Key features: - **Data Curation**: Scrapes data from arXiv, GitHub, StackOverflow, HuggingFace, and custom URLs, cleans it using MinHash LSH and PII removal, and filters based on quality metrics. - **Training**: Uses QLoRA + DPO training, DeepSpeed ZeRO-3, Flash Attention, and HuggingFace-compatible models. - **Model Merging**: Combines models using TIES, DARE, and SLERP methods, with weights optimized for performance. - **Evaluation**: Evaluates models using benchmarks like MMLU, LLM-as-judge, and statistical significance. - **Deployment**: Deployed using Docker and AWS spot instances, with FastAPI and Terraform configurations. ### Tech stack: Python, DeepSpeed ZeRO-3, Flash Attention, Prometheus, Docker, AWS spot instances, FastAPI, Terraform. ### Why it matters: This pipeline is designed to streamline the fine-tuning, merging, and deployment of large language models, making it highly adaptable and efficient for various HuggingFace-compatible models.
Python 5 stars

TerboucheHacene/speech-keyword-spotting

Speech Keyword detection using Wav2Vec Model

Overview: This project implements a speech keyword spotting system using a pretrained Wav2Vec2 model, providing a custom model for detecting keywords based on raw audio. Key features: - Custom model training on a pretrained Wav2Vec2 model - Evaluation and experiment tracking using Weights & Biases - Exporting the model to TorchScript and ONNX formats - Deploying the model as a REST API using FastAPI - Containerizing the code using Docker and NGINX as a web proxy server Tech stack: Python Why it matters: This project showcases the effectiveness of using pretrained models for speech recognition tasks, demonstrating how limited labeled data can still lead to significant improvements in accuracy.
Python 5 stars

Thatayotlhe04/Parallelogram

Linter for LLM fine-tuning datasets: catches broken chats and tool calls, context overflow, leaked keys, PII and eval contamination before you train. CLI, Python API, MCP server, GitHub Action.

### Overview: A strict CLI validator that hard-blocks on bad data before any compute is spent. ### Key features: - **Data Validation**: The validator checks for and blocks any data that could corrupt training, ensuring clean and reliable fine-tuning. - **CLI Interface**: The tool is designed as a command-line interface (CLI) for easy integration into workflows. - **Agent Support**: The CLI includes support for various fine-tuning agents, including MCP tools, Claude Code/Codex/Cursor configuration, Python errors, and a six-line AGENTS.md/CLAUDE.md snippet guide. ### Tech stack: Python ### Why it matters: Ensures that fine-tuning runs are not interrupted by corrupted data, maintaining the integrity and quality of the training process.
Python 5 stars

TirendazAcademy/Lightly-Train-Tutorials

This repo shows how to use the LightlyTrain framework

Overview: This repository showcases how to use the LightlyTrain framework for training computer vision models on unlabeled data. Key features: - Works with raw image folders (no annotations required) - Supports pretrained backbones like ResNet - Easily fine-tune for classification later Tech stack: Python Why it matters: LightlyTrain simplifies the process of training computer vision models on unlabeled data, making it accessible and efficient for beginners and advanced users alike. GitHub stars: 5 Repository: TirendazAcademy/Lightly-Train-Tutorials
Python 5 stars

Traversecityjesuit102/open_qwen

Build and deploy large language models with Open Qwen, a PyTorch-based framework for efficient, flexible AI development and research.

Overview: Open Qwen is a user-friendly application based on the Qwen 3.5 model, designed to provide AI-powered text generation without the need for technical setup or programming skills. It focuses on practical tasks and is accessible on Windows. Key features: - Text generation based on AI technology - Supports questions and answers - Lightweight and easy to install - Runs locally on Windows computers - No programming needed to use - Built with common AI tools like PyTorch Tech stack: Python Why it matters: Open Qwen simplifies the process of using AI for text tasks, making it accessible to a broader audience without requiring advanced technical knowledge.
Python 5 stars

ZLHAOOO/laya-mlx-zh

Chinese fine-tuned Laya decision model weights (MLX, Apple Silicon) — 中文微调 Laya 决策模型: 消息路由/紧急度/优先级/检索选词, 一次前向全出, ~27ms

Overview: This repository contains the fine-tuned Laya decision model weights (MLX, Apple Silicon) for Chinese language processing tasks, specifically focusing on message routing, urgency assessment, priority ranking, and retrieval selection. The model achieves an accuracy of 0.85-0.90 in Chinese decision-making tasks, making it highly relevant for developers working with Apple Silicon devices. Key features: - Fine-tuned Laya decision model weights for Chinese language processing tasks. - Supports Apple Silicon devices. - Achieves an accuracy of 0.85-0.90 in Chinese decision-making tasks. - Includes a demonstration of how to handle new scenarios without retraining. Tech stack: Python Why it matters: This model is highly relevant for developers working with Apple Silicon devices, offering a fine-tuned solution for Chinese language processing tasks that can be deployed without the need for GPU, API subscriptions, or real-time performance issues. Repository: ZLHAOOO/laya-mlx-zh GitHub stars: 5 GitHub description: Chinese fine-tuned Laya decision model weights (MLX, Apple Silicon) — 中文微调 Laya 决策模型: 消息路由/紧急度/优先级/检索选词, 一次前向全出, ~27ms
Jupyter Notebook 5 stars

ZahraSahranavard/Gender-Image-Classification

Overview: This project aims to classify images of men and women using Deep Learning techniques, specifically leveraging Convolutional Neural Networks (CNNs) and pre-trained VGG16 architecture. Key features: - Fine-tuning of VGG16 architecture on a custom dataset - Classification of images into male and female categories - Utilization of pre-trained CNNs for improved accuracy on small datasets - Data augmentation for robustness Tech stack: Python, TensorFlow, Keras, NumPy, pandas, Matplotlib Why it matters: This project showcases the effectiveness of transfer learning and fine-tuning in improving model performance on small datasets, demonstrating how pre-trained models can be fine-tuned to achieve high accuracy.