Projects

Overview: Train Studio is a web-based user interface designed for training Low-Rank Fusion (LoRA) models using SFT (Scaling Fusion Training) on GPU hardware. It allows users to select GPU, model, dataset, and configuration parameters directly through the browser, providing a seamless experience akin to unsloth.ai. Key features: - Hardware detection and selection - Model and dataset configuration - Training configuration management - Real-time training status and log refresh - Automatic LoRA merge and full model generation Tech stack: Python, Gradio, Hugging Face Transformers, PyTorch, CUDA, GPU, GPU memory management, and GPU-specific libraries like nvidia-smi and /proc. Why it matters: Train Studio simplifies the process of training LoRA models, making it accessible to developers and researchers without requiring deep expertise in GPU management or complex training scripts. It streamlines the workflow by automating common tasks and providing a user-friendly interface, enhancing productivity and reducing the learning curve for new users.

### Overview: DataForge is a comprehensive and user-friendly platform designed specifically for fine-tuning Large Language Models (LLMs) datasets. It seamlessly integrates with your existing data sources and offers a streamlined workflow for generating, editing, and exporting datasets tailored to modern LLMs. ### Key Features: - **Import**: Supports a wide variety of data formats including JSONL, JSON, CSV, Parquet, Excel, PDF, DOCX, Markdown, and plain text. It also imports datasets directly from Hugging Face Hub. - **Dataset Types**: Manages four types of datasets: SFT chat conversations, DPO preference pairs, KTO unpaired feedback, and RL prompts. - **Reasoning Traces and Tool Calls**: Keeps these data fields separate to ensure correct rendering for each target model. - **Edit**: Features a virtualized grid that handles large datasets without performance issues, including role switching, reasoning blocks, tool call cards, and per-message loss masking. - **Quality Checks**: Includes 17 issue checks for empty messages, broken tool calls, PII, encoding damage, and context overflow. - **Generate**: Offers synthetic data generation, Q&A from documents, and on-policy DPO pair building. - **Export**: Provides various export options for compatibility with different LLMs, including JSONL, Axolotl 0.17, TRL 1.5, LLaMA-Factory 0.9.5, MS-SWIFT 4.3, Unsloth, and OpenAI fine-tuning. - **Model Registry**: Tracks

### Overview: These projects showcase various computer vision tasks using PyTorch, including image classification, object detection, and text detection. The author fine-tunes pre-trained models and exports them to ONNX format for easy deployment. ### Key Features: - **Image Classification**: Fine-tuned ResNet-50 and Google VIT-Base-Patch16-224 for cats and dogs, and plant species respectively. - **Object Detection**: Fine-tuned Faster-RCNN and YOLO-V10 for text detection and facemask detection, respectively. - **Text Detection**: Fine-tuned Faster-RCNN for manga text detection. ### Tech Stack: Python ### Why it Matters: These projects demonstrate practical applications of PyTorch in computer vision tasks, showcasing the effectiveness of fine-tuning pre-trained models and the importance of model export to ONNX format for deployment.

Overview: This project involves fine-tuning a ResNet-18 model on the CIFAR-100 dataset using transfer learning, aiming to improve classification accuracy. Key features: - Fine-tuning a ResNet-18 model on CIFAR-100 - Comparison of pretrained models - Fine-tuning the Resnet-18 model on CIFAR-100 - Using augmentations to improve robustness - Testing the model for robustness to Gaussian noise Tech stack: Jupyter Notebook Why it matters: This project demonstrates effective transfer learning techniques and showcases the potential of fine-tuning pre-trained models for improved classification accuracy on new datasets.

Overview: This project involves developing a predictive model to forecast the number of online shares an article will receive, using various attributes of the article as input variables. Key features: - Data preprocessing techniques for handling missing values and normalizing data. - Feature engineering to select relevant variables for the model. - Model training and evaluation using machine learning algorithms. Tech stack: Python Why it matters: Understanding and predicting the popularity of online news articles can help in optimizing content strategies, improving engagement, and enhancing the overall reach of news platforms. Repository: north0n-FI/Predicting-the-Popularity-of-Online-News Primary language: Jupyter Notebook GitHub stars: 5

Overview: This repository demonstrates fine-tuning GPT-2 on the Stanford Question Answering Dataset (SQuAD) using the Hugging Face distilgpt2 model. Key features: - Fine-tuning GPT-2 on SQuAD - Using Hugging Face distilgpt2 for model adaptation - Evaluating results with perplexity Tech stack: Jupyter Notebook Why it matters: This project showcases the practical application of fine-tuning techniques in natural language processing, demonstrating how pre-trained models can be adapted for specific tasks.

Overview: Onde Inference CLI is a command-line interface for managing and fine-tuning Qwen models on Apple silicon devices, including macOS. Key features: - Fine-tune Qwen models locally with LoRA - Export models to GGUF format - Manage account, apps, and model catalog - Terminal UI for local management - MCP server for remote management Tech stack: Rust Why it matters: Onde Inference CLI simplifies the process of fine-tuning and managing Qwen models, making it accessible on Apple silicon devices and allowing for remote management through MCP servers. Repository: ondeinference/onde-cli GitHub stars: 5 GitHub description: Fine-tune, quantize, and publish your model all from your Apple silicon devices.

Overview: This project demonstrates how to serve a trained `doddle-model` in a pipeline implemented with Apache Beam and Scio, utilizing batching for faster vectorized predictions. Key features: - Serving a trained `doddle-model` in a Beam pipeline. - Batch processing of examples for improved performance. Tech stack: Scala, Apache Beam, Scio. Why it matters: This example showcases the integration of a machine learning model with a data processing pipeline, highlighting the benefits of using Apache Beam for efficient data handling and processing. Repository: `picnicml/doddle-beam-example` Primary language: Scala GitHub stars: 5 GitHub description: An example of serving a `doddle-model` in a Beam pipeline, including batching for faster predictions. README excerpt: An example of how to serve a trained `doddle-model` in a pipeline implemented with Apache Beam and Scio, including batching for faster vectorized predictions.

### Overview: FreezeAsGuard is a technique designed to mitigate illegal adaptations of diffusion models by selectively freezing tensors critical for adaptation but preserving their representation power on innocent domains. This approach is essential for maintaining the integrity of diffusion models and ensuring their utility across various applications. ### Key Features: - **Selective Tensor Freezing**: The method selectively freezes model tensors that are adaptation-critical for illegal domains but retain representation power on innocent domains. - **Pretrained Model**: The technique utilizes a pretrained model to mitigate illegal adaptations. - **Dataset Preparation**: The dataset `FamousFigures-25 (FF25)` is downloaded and prepared for use with the model. ### Tech Stack: Python ### Why It Matters: FreezeAsGuard is crucial for maintaining the security and integrity of diffusion models, ensuring they remain effective and reliable across different applications without being exploited for illegal purposes.

```

### Overview: This dataset contains 12,000 examples across 42 categories for cyber threat intelligence tasks, designed specifically for instruction-tuning fine-tuning of large language models (LLMs). It includes MITRE ATT&CK, CVE/KEV, CWE, CAPEC, D3FEND, and live threat feeds, ensuring high accuracy and reliability. ### Key Features: - **Categories**: 42 distinct categories with at least 155 examples each. - **Entity Coverage**: Each record is grounded and covers the same entity across multiple tasks, ensuring consistency. - **Distinct Instructions**: 95.8% of instructions are distinct, providing a diverse range of tasks. - **Split Strategy**: Train and validation splits share no entities, ensuring each task is isolated. - **Format**: Single-turn records use `instruction`, `input`, and `output`, while multi-turn records are in `messages` format. ### Tech Stack: - **Data Sources**: MITRE ATT&CK, CVE/KEV, CWE, CAPEC, D3FEND, MITRE Engage, CISA KEV, CISA CSAF ICS advisories, OSV.dev, SigmaHQ, abuse.ch (Feodo, SSLBL, URLhaus, ThreatFox), MalwareBazaar, Malpedia, ransomware.live, OpenPhish, OWASP Top 10, API Top 10 and Top 10 for Agentic Applications 2026, STRIDE, NIST SP 800-61,

Overview: This project is an OpenAI-compatible FastAPI server designed for German neurology and neuro-oncology audio transcription using Parakeet-MLX models. It is optimized for Apple Silicon M4 and compatible with M1/M2/M3/M4 devices. Key features: - OpenAI-compatible API - Apple Silicon optimized for M4 - Supports various audio formats (WAV, MP3, FLAC, M4A, etc.) - Optimized for German neurology and neuro-oncology terminology - Multilingual support - Fast inference with MLX - Easy to use installation - Built-in Swagger UI and ReDoc for interactive API docs - Health monitoring endpoint for production use Tech stack: Python Why it matters: This server is specifically designed to work with the NeurologyAI/neuro-parakeet-mlx model, optimized for German neurology and neuro-oncology terminology, and compatible with Apple Silicon devices, making it ideal for speech recognition tasks in the healthcare industry.

Overview: Dermagemma is an offline-first dermatology AI solution designed for skin of color, addressing the systemic bias and deployment challenges faced by AI tools in regions with limited internet access and infrastructure. It combines a fine-tuned Vision Transformer classifier, a hybrid RAG retriever, local Gemma 4 E2B-IT model, and WeasyPrint for a clean, downloadable PDF clinical chart, all running on a CPU laptop with no network connection required after setup. Key features: - Fine-tuned Vision Transformer classifier trained with Skin of Color emphasis - Hybrid RAG retriever with explicit Skin of Color tagging and re-ranking boost - Local Gemma 4 E2B-IT model running via llama.cpp - WeasyPrint for clean, downloadable PDF clinical chart Tech stack: Python, llama.cpp, BM25, PubMedBERT, DermNet NZ knowledge base, WeasyPrint Why it matters: Dermagemma addresses the systemic bias and deployment challenges in dermatology AI, providing accurate and accessible dermatological assessment for skin of color, particularly in regions with limited internet access and infrastructure.

Overview: This repository provides a fine-tuned version of the BERT base-uncased model orchestrated with Flyte and deployed on Banana, enabling efficient model development and inference. Key features: - Flyte workflow orchestration platform for data and ML pipelines. - Banana for inference hosting on serverless GPUs. - HuggingFace for seamless model storage and retrieval. Tech stack: Python Why it matters: This setup allows for streamlined model development and deployment, leveraging the benefits of Flyte for orchestration and Banana for efficient model inference, making it ideal for developers looking to streamline their ML pipeline workflows.

Overview: Moleculegen-ML is a Python package designed for de novo drug design using generative language modeling, focusing on novel molecule generation. Key features: - Implements tools for molecular data processing - Utilizes SMILES-based language modeling techniques - Supports recurrent networks, autoregressive transformers, and convolutional networks - Supports transfer learning Tech stack: Python Why it matters: This package offers a novel approach to drug design by leveraging generative language models, potentially speeding up the process and expanding the scope of possible molecule designs.

Overview: This project fine-tunes StarCoder2-3b for SQL tasks using LORA, reducing model size for faster training on smaller datasets. Key features: - Fine-tuning StarCoder2-3b for SQL tasks - Utilizing LORA for reduced model size - Training on 600+ programming languages and natural language text Tech stack: Jupyter Notebook Why it matters: This approach optimizes performance and resource efficiency, making it suitable for applications requiring limited computational resources.

Overview: This project involves finetuning a Large Language Model (LLM) to extract structured data from press releases, leveraging a previously annotated dataset from the ISAF Press Releases project. Key features: - Finetuning an LLM for structured data extraction - Leveraging a pre-existing dataset for structured data extraction - Using the ISAF Press Releases dataset for context and annotation Tech stack: Jupyter Notebook Why it matters: This project offers an interesting use case for structured data extraction, providing a practical application for the finetuning of an LLM.

Overview: This project aims to analyze financial news articles and predict their sentiment using a pre-trained sentiment analysis model, focusing on market trends and investor sentiment for stock trading and investments. Key features: - Sentiment Analysis: Predicts the sentiment of financial news articles. - Model Integration: Utilizes a fine-tuned DistilRoBERTa model optimized for financial sentiment analysis. - Dataset Support: Processes financial datasets for model training and evaluation. Tech stack: Python-based project using Jupyter Notebook. Why it matters: This project helps investors and traders understand market trends and investor sentiment, aiding informed decision-making in stock trading and investments.

Overview: This repository contains code and documentation for analyzing the multi-task FairFace dataset and building/fine-tuning various models like Mobilenet, ResNet, and implementing contrastive learning techniques. Key features: - Analysis of the FairFace dataset - Building and fine-tuning models - Implementing contrastive learning techniques Tech stack: Jupyter Notebook Why it matters: This project provides a comprehensive solution for facial attribute classification using the FairFace dataset, demonstrating the effectiveness of modern deep learning techniques in real-world applications.

Overview: faster_outlines is a Rust backend designed to significantly enhance the performance of regex-guided text generation, particularly for LLM inference servers. Key features: - Seamless one-line integration with existing Outlines projects - All the features of the Outlines library - Asynchronous FSM compilation for immediate LLM inference start - Substantial performance improvements, especially for complex regex patterns - Continuous updates for speed improvements Tech stack: Rust Why it matters: faster_outlines optimizes performance for LLM inference servers, making it ideal for dynamic regex patterns and complex state spaces. Repository: unaidedelf8777/faster-outlines GitHub stars: 5 GitHub description: A Lazy, high throughput and blazing fast structured text generation backend.