Project directory

Projects

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

Python 5 stars

a-sydelov/ai-service-skeleton

Practical skeleton for ML/AI services with controlled rollout, primary/shadow serving, observability, and cutover/rollback.

Overview: This repository provides a practical skeleton for developing ML/AI services, focusing on model serving and controlled rollout. It includes primary/shadow serving, traffic splitting, shadow mirroring, cutover, and rollback functionalities. Key features: - Primary/shadow model versions - Traffic split - Shadow mirroring - Cutover and rollback - Canary-style rollout Tech stack: Python Why it matters: This skeleton serves as a foundational framework for extending and adapting ML/AI services to various models, serving patterns, and operational environments, making it easier to manage and validate model promotions in a live service.

Python 5 stars

abdalaziz-saif/end-to-end-vehicle-classifier

End-to-end vehicle type classification pipeline using transfer learning, trained on web-scraped images (Bing/Baidu/Googel) with automated data cleaning, validation, and preprocessing.Achieves 89.7% validation accuracy across 9 classes

Overview: This project automates the classification of vehicle types from images using a deep learning model trained on web-scraped images. It achieves 89.7% validation accuracy across 9 classes, including cars, motorcycles, buses, trucks, and bicycles. Key features: - Web scraping for image data collection - Transfer learning with MobileNetV2 - Advanced data augmentation techniques - Fine-tuning strategy with learning rate scheduling - Early stopping to prevent overfitting - Data cleaning and validation - Support for 9 vehicle types Tech stack: Python Why it matters: This project demonstrates the effectiveness of deep learning models in real-world applications, particularly in automated image classification tasks. Repository: abdalaziz-saif/end-to-end-vehicle-classifier GitHub stars: 5 GitHub description: End-to-end vehicle type classification pipeline using transfer learning, trained on web-scraped images (Bing/Baidu/Googel) with automated data cleaning, validation, and preprocessing. Achieves 89.7% validation accuracy across 9 classes.

Python 5 stars

abgache/Genmoji

Reproduction of Apple Intelligence Genmoji for apple style emoji generation in python using PyTorch, it's a fine-tune of Stable Diffusion 1.5.

Overview: This project is a fine-tuned version of Stable Diffusion 1.5, specifically designed for generating Apple-style emoji in Python using PyTorch. Key features: - Reproduction of Apple Intelligence Genmoji - Works with NVIDIA GPUs - Supports training and generating GenMoji - Enhances prompts with detailed textures and 3D lighting - Generates detailed and enlarged emoji representations Tech stack: Python, PyTorch, Stable Diffusion 1.5 Why it matters: This project offers a unique way to generate Apple-style emojis, showcasing the versatility of AI in creating stylized representations of popular characters.

Jupyter Notebook 5 stars

adamlawal/friendlies

A causal language model trained to speak like colleagues of mine

Overview: This project is a causal language model trained to mimic the conversational style of the author's friends, using a custom tokenizer and a small Llama-style decoder architecture. Key features: - Custom byte-level BPE tokenizer (~8k vocab) from the GPT-2 template - Tiny Llama-style decoder architecture - Messages grouped by speaker and hour, with a per-person train/val split - Generate with prompts like `Koded:` or `Adam:` or any of your friends names if you try it out 🗣️ Tech stack: Jupyter Notebook Why it matters: This project demonstrates the capabilities of causal language modeling by creating a conversational AI that speaks like a group of friends, showcasing the potential of training models on real-world conversations. Repository: adamlawal/friendlies GitHub stars: 5 GitHub description: A causal language model trained to speak like colleagues of mine

Python 5 stars

aifeifei798/unsloth-gui

Lightweight local web UI for fine-tuning LLMs with Unsloth on a single GPU (FastAPI + Jinja2). Data prep, LoRA SFT training with resume and live TensorBoard, streaming chat testing. Bilingual EN/ZH UI, text SFT only.

Overview: This project is a lightweight local web UI for fine-tuning LLMs with Unsloth, focusing on pure text single-GPU SFT training. It includes data preparation, LoRA SFT training with resume and live TensorBoard, streaming chat testing, and bilingual EN/ZH UI with text SFT only. Key features: - Data processing - LoRA SFT training with resume and live TensorBoard - Streaming chat testing - Model management (local, HF, and Magician) - Data management (unified data list, renaming, deletion) - Interface language setting Tech stack: FastAPI + Jinja2 + Unsloth Why it matters: This project provides a user-friendly interface for fine-tuning large language models using Unsloth, making the process of fine-tuning LLMs more accessible and efficient. The web UI allows users to manage models, data, and interface language settings without needing to write code or modify configuration files. The integrated data processing and live monitoring features streamline the training process, ensuring that the model and interface language are unified and easily managed.

Python 5 stars

alireza13742010/LIAMA3_chatbot_Arabic_Engligh

This repository is for developing the chatbot on health care question and answer

Overview: This repository is for developing a chatbot that can answer health care questions in multiple languages, including Arabic, Persian, English, and Spanish. Key features: - Supports multiple languages: Arabic, Persian, English, and Spanish. - Uses a large pre-trained model with 7 billion parameters. - Trained on 20 million questions and answers. - Available on GitHub with 5 stars. Tech stack: Python. Why it matters: This project showcases the potential of large language models in healthcare, providing accurate and multilingual answers to health-related queries. Repository: alireza13742010/LIAMA3_chatbot_Arabic_Engligh Primary language: Python GitHub stars: 5

Jupyter Notebook 5 stars

alpcansoydas/turkish-hate-speech-analysis

Turkish Hate Speech Detection

Overview: This project implements a Turkish hate speech recognition system using a fine-tuned version of YSKartal/bert-base-turkish-cased-turkish_offensive_trained_model and dbmdz/bert-base-turkish-cased. The dataset consists of labeled data from the Turkish Offensive Language Detection dataset. Key features: - Fine-tuned hate speech detection model - Uses a pre-trained model for improved performance - Labeled dataset for training and evaluation Tech stack: Python Why it matters: This project is significant for its contribution to the field of hate speech detection in Turkish, which is crucial for maintaining a safe and inclusive digital environment.

Python 5 stars

amjadmajid/llm_toaster

LLM Toaster enables you to train and fine-tune mini-GPTs.

### Overview LLM Toaster is a library designed for training, fine-tuning, and running inference on Transformer-based language models. It streamlines the process of dataset loading, model configuration, and execution. ### Key features: - **Model Configuration**: Configurable model architecture with options for depth, width, and heads. - **Training and Fine-Tuning**: Supports training on various datasets and fine-tuning on pre-trained models. - **Inference**: Provides efficient inference capabilities for models trained with LLM Toaster. ### Tech stack: Python ### Why it matters: LLM Toaster simplifies the process of training and fine-tuning large language models, making it easier for developers to experiment with different architectures and configurations.

Python 5 stars

arkodeepsen/helix

Professional training stack for 100M parameter language models optimized for AMD Radeon RX 7000/9000 series GPUs with ROCm. Complete pipeline: tokenization, pretraining, fine-tuning, DPO, and GGUF export.

Overview: Helix is a complete training stack designed for building small language models optimized for AMD GPUs using ROCm. It includes tokenization, pretraining, fine-tuning, data preparation, and export to GGUF format. Key features: - Tokenization and pretraining - Fine-tuning with supervised fine-tuning (SFT) - Data preparation and export to GGUF format - Agent demonstration with tool calls - Hugging Face deployment Tech stack: Python Why it matters: Helix is optimized for AMD GPUs, making it suitable for high-performance language model training on modern graphics cards.

Jupyter Notebook 5 stars

aryanap07/Timeless

Before the prompt, there was the architecture.

Overview: Timeless is a project that involves building a large language model from scratch, inspired by the architecture described in "Building a Large Language Model (From Scratch)" by Sebastian Raschka. Key features: - Large Language Model from Scratch - Architecture Reference Tech stack: Jupyter Notebook Why it matters: Timeless serves as a practical demonstration and educational resource for understanding the foundational principles of large language models, providing insights into the architecture and implementation of such models.

Jupyter Notebook 5 stars

athlohangade/qa-system-using-tl-legal-domain

A Question Answering System built for answering specific questions on Indian Legal Acts following the Transfer Learning approach.

Overview: This repository is a Question Answering System (QAS) designed to answer specific questions on Indian Legal Acts using a Transfer Learning approach. The system focuses on legal documents, particularly those amended by the Indian Constitution and Judiciary, and is built on Jupyter Notebook. Key features: - Fine-tuned BERT-Large model for legal domain dataset - Preprocessing and context creation - Question generation using Google T5-base - Batch processing with a batch size of 32 for question answering Tech stack: Jupyter Notebook Why it matters: This QAS system provides a practical application of Transfer Learning in the field of NLP, specifically for legal documents, enhancing efficiency and reducing domain-specific dataset size.

Python 5 stars

aws-samples/sample-vla-mobile-manipulation-on-aws

Bring-your-own VLA robot learning on AWS: custom task spec, teleop data collection, Bedrock annotation, openpi/OpenVLA-OFT fine-tuning on EC2, rolling-window evaluation, and phase-routed multi-model deployment for a quadruped + 6-DoF arm

Overview: This sample demonstrates the process of bringing your own VLA stack on AWS, focusing on a custom task specification, teleop data collection, Bedrock annotation, fine-tuning on EC2, rolling-window evaluation, and phase-routed multi-model deployment for a quadruped + 6-DoF arm robot. Key features: - Custom four-phase mobile pick-and-place task - Self-built leader-follower teleop rig with dual cameras and operator keypress phase labels - Custom locked-vocabulary schema and multi-frame voting on Amazon Bedrock - Forked openpi: 10-D action fusing locomotion + manipulation, 7-D state, dual-camera input - Rolling-window EE-error protocol (mean / k=0 / k=last) over 30+ checkpoints - Phase-Routed VLA (PR-VLA): route each phase to the model that earns it, async chunk client, on-device safety guards Tech stack: Python Why it matters: This sample provides a comprehensive guide on how to customize every layer of the VLA stack on AWS, including task definition, data collection, annotation, model configuration, evaluation, and deployment, making it a valuable resource for developers looking to build their own VLA stack on the platform. GitHub stars: 5 Repository: aws-samples/sample-vla-mobile-manipulation-on-aws Primary language: Python Architecture: ![Architecture](assets/architecture.png) How this differs from other AWS samples: This sample focuses on customizing every layer of the VLA stack, including task definition

Python 5 stars

blackccpie/GweLLM

Journey towards Fine-Tuning a Breton speaking Chat Model

### Overview: A developer-focused project aimed at fine-tuning a Breton speaking Chat Model. The project began with an initial motivation to create open-source lightweight language models for Breton, focusing on local deployment and ease of use. The primary language used is Python, and the project has gained 5 GitHub stars. ### Key Features: - Fine-tuning of a pretrained Conversational LLM to a new language, specifically Breton. - Development of a French -> Breton translation model called **Gallek**. - Creation of a Breton Instruction Dataset called **Goulenn**. - Development of a Breton Conversational LLM called **GweLLM**. ### Tech Stack: - Python ### Why It Matters: This project aims to bridge the gap between Breton language and technology, making it easier for local developers and enthusiasts to interact with Breton through conversational models. By providing a practical tool for language translation and instruction dataset creation, it democratizes the use of Breton in everyday applications.

Python 5 stars

caiotheodoro/specula

Fine-tune a VL Qwen on FDA food-label compliance: image in, cited PASS|FLAG out. Seeded synthetic labels + CFR verifier-as-oracle + SFT/GRPO.

### Overview: Fine-tune a multimodal Qwen on FDA food-label compliance: image in, cited PASS|FLAG out. This project fine-tunes a 27B Qwen model on FDA regulations to improve compliance with 21 CFR Part 101 and Part 104, using a deterministic oracle for verification. The model is trained on free credits on Modal/GCP, making it accessible and cost-effective. ### Key features: - Multimodal fine-tuning of a Qwen model. - Training on FDA regulations, including 21 CFR Part 101, Part 104, FALCPA allergen rules, and 2026 "healthy" and FOP rules. - Use of a deterministic oracle for verification, ensuring exact flag computation. - Full process automation, including intake, rule application, verdict, and report generation. - Benchmark contracts for fixed taxonomy, scoring functions, and contamination scheme. ### Tech stack: Python ### Why it matters: This project leverages multimodal fine-tuning to improve regulatory compliance, making it approachable and cost-effective. The deterministic oracle ensures that the model's performance is aligned with regulatory standards, providing a high level of accuracy and reliability.

Python 5 stars

caiotheodoro/suture

Fine-tune a VL Qwen on insurance policy-issuance QC: binder vs issued policy images in, structured FLAG/PASS out. Seeded synthetic + verifier-as-oracle + SFT/ReST-EM.

### Overview: Fine-tune a VL Qwen model on insurance policy-issuance QC: compare binder and issued policy images, generating a structured flag/pass output. The model is trained on synthetic data with a verifier-as-oracle and uses structured JSON predictions. ### Key Features: - **Model:** Fine-tuned Qwen/Qwen3-VL-8B-Instruct model. - **Dataset:** Seeded synthetic data for training. - **Oracle:** Verifier-as-oracle for high accuracy. - **Output:** Structured JSON with PASS/FLAG and discrepancy list. - **Recall/Precision:** 0.959, 0.969. ### Tech Stack: - **Model:** Qwen/Qwen3-VL-8B-Instruct - **Dataset:** Seeded synthetic data - **Oracle:** Verifier-as-oracle - **Adapter:** QLoRA on Qwen/Qwen3-VL-8B-Instruct ### Why It Matters: Policy issuance QC is a closed, finite ontology task with exact ground truth. The combination of finite ontology, exact ground truth, and no dedicated model makes it solvable with a small adapter. The same pattern used on ReconForge also works for this task, demonstrating the power of small adapters in closed tasks. ### Result: - **Recall/Precision:** 0.959/0.969 - **High Recall/Precision:** 0.969 - **Deductible Direction:** Closed from 0.507/0.

TypeScript 5 stars

cedrickdev/NeoCraft-Agency-Landing

### Overview: Neocraft Agency Landing est un site professionnel développé avec Next.js 16, React 19, TypeScript, et Tailwind CSS. Il est conçu pour l'agence NeoCraft basée à Lausanne, Suisse, et comprend un système internationalisé, un blog géré par Wisp CMS, et une gestion des erreurs via Sentry. ### Key features: - Next.js 16 pour l'application front-end - React 19 pour l'interface utilisateur - TypeScript pour le type statique - Tailwind CSS pour les styles - Internationalisation via next-intl - Envoi d'emails transactionnels via Resend - Gestion du blog via Wisp CMS - Animations via Framer Motion - Composants UI via shadcn/ui ### Tech stack: Next.js 16, React 19, TypeScript, Tailwind CSS, next-intl, Sentry, Resend, Wisp CMS, Framer Motion, shadcn/ui ### Why it matters: Cette application est un site professionnel conçu pour l'agence NeoCraft, offrant une interface utilisateur moderne, une gestion internationalisée, et un système de monitoring d'erreurs pour assurer la fiabilité du site.

Python 5 stars

chaeminyoon/AIS-Traffic-Ops

An MLOps project for forecasting maritime traffic patterns from AIS data and operating the model through evaluation, serving, and monitoring workflows.

### Overview: AIS Traffic Ops is an MLOps project designed to forecast maritime traffic patterns from AIS data. It preprocesses AIS inflow/outflow arrays into 66x66 grid sequences, experiments with various architectures, tracks model versions, manages the active production model through a `production` symlink, supports hot reload for model replacement, and provides serving, monitoring, model registry, inference logs, and dashboard views. ### Key Features: - Preprocesses AIS inflow/outflow arrays into 66x66 grid sequences. - Experiments with ConvLSTM, temporal attention, multi-head attention, spatial UNet, and gated UNet variants. - Tracks model versions from `v1` through `v16`. - Manages the active production model through a `production` symlink. - Supports hot reload for model replacement without restarting the API server. - Provides serving, monitoring, model registry, inference logs, and dashboard views. ### Tech Stack: - Python 3.10+ - FastAPI serving - Prometheus and Grafana monitoring - MLflow model registry - SQLite inference logs - Custom dashboard ### Why It Matters: This project is crucial for maritime traffic forecasting, enabling more accurate predictions and improved operational efficiency. The use of various architectures and the ability to track and manage model versions ensure that the system remains up-to-date and effective. The integration with Prometheus, Grafana, and MLflow enhances monitoring and reproducibility, making it easier to understand and reproduce the system structure.

Jupyter Notebook 5 stars

chriss1245/mlflow_packaging_samples

Worked examples of packaging PyTorch models with MLflow — native flavour vs custom pyfunc, and what each costs at serving time.

Overview: This repository provides examples of packaging PyTorch models with MLflow, showcasing the differences between the native flavor and a custom `pyfunc` approach. It highlights the costs associated with each choice at serving time. Key features: - Native flavor with `mlflow.pytorch` - Custom `pyfunc` with `mlflow.pyfunc` Tech stack: Jupyter Notebook Why it matters: Understanding the trade-offs between these packaging methods is crucial for optimizing the performance and cost of serving models, especially when dealing with complex preprocessing and postprocessing requirements.

Python 5 stars

daekeun-ml/encoder-serving-sagemaker

Benchmarked comparison of multiple serving patterns for Transformer encoder models (BERT-style classification, NLI) on Amazon SageMaker AI — ONNX CPU, multi-container, inference component, LMI, scale-to-zero

Overview: This repository provides a comprehensive guide on deploying Transformer encoder models, such as BERT-style classification and NLI, on Amazon SageMaker AI. It benchmarks multiple serving patterns to determine the most efficient and cost-effective method for real-time endpoint deployment. Key features: - Benchmarks multiple serving patterns for Transformer encoder models. - Compares latency, throughput, and cost for different deployment scenarios. - Includes real-time endpoint deletion warning. Tech stack: Python Why it matters: This repository helps developers understand the best practices for deploying Transformer encoder models on SageMaker AI, considering factors like latency, throughput, and cost efficiency.

Python 5 stars

daekeun-ml/frontier-moe-serving-benchmark

Serving and benchmarking 9 frontier MoE checkpoints (490 GiB+) with vLLM on a single 8x B200 node, plus a pre-download HBM fit checker

### Overview: This repository serves and benchmarks nine frontier MoE checkpoints, totaling 490 GiB or larger, across eight B200 GPUs. It provides comprehensive serving configurations, benchmark results, and detailed documentation for optimal deployment on a single node. ### Key Features: - Serving configurations for nine frontier MoE checkpoints. - Benchmark results across five scenarios (A.X-K2 supports uv only). - Detailed documentation and reproducibility with one command. - Measured performance and memory usage across multiple checkpoints. - Optimal GPU count and memory estimation for different hardware configurations. ### Tech Stack: Python ### Why It Matters: This repository provides essential information for deploying frontier MoE models on a single node, ensuring accurate performance and memory usage predictions.