Projects

Overview: tLLM is a Python library designed to simplify the fine-tuning of large language models (LLMs) for instruction-tuning on custom datasets. It abstracts away complexities from the Hugging Face ecosystem, making the process more accessible and user-friendly. Key features: - Fine-tunes LLMs using HF's `Trainer` for custom datasets stored in the Hub. - Loads model with 4-bit quantization. - Uses LoRA under the hood, combining 4-bit quantization to reduce memory requirements. - Converts dataset into a prompt template for fine-tuning. - Tracks and logs experiments using W&B. Tech stack: Python Why it matters: tLLM simplifies the process of fine-tuning LLMs, making it more accessible, especially for beginners in the realm of AI. It provides a more encapsulated and user-friendly approach, lowering the barrier of entry using the state-of-the-art open-source stack. GitHub stars: 26

Overview: JAiRouter is a production-ready AI model gateway that simplifies managing multiple LLM backends by providing a single, OpenAI-compatible API endpoint. Key features: - Unified API surface for multiple LLM backends - Load balancing - Rate limiting - Circuit breaking - Failover to a different healthy instance - RBAC (Role-Based Access Control) - Audit logging - Centralized metrics and tracing Tech stack: Java 17 + Spring Boot 3.5.5 (WebFlux / reactive) Why it matters: By centralizing the management of multiple LLM backends under a single API, JAiRouter allows developers to focus on building applications rather than managing infrastructure. This simplifies the deployment and scaling of AI models, making it easier to manage and scale the backend services without the need for manual failover or complex configuration changes.

Overview: JustRayzist is a lightweight CUDA-based local inference platform built around Z-Image Turbo by Tongyi, designed for fine-tuning and image generation. It offers a seamless user experience with a fast, stable backend, and supports a variety of features including prompt enhancement, x2 SeedVR2 upscale, and more. Key features: - Local web UI, API, and CLI in one repository - Auto resource-tier detection with stable normal-user defaults - Model-pack runtime with public, hidden, disabled, and derived runtime pack behavior - Multi-LoRA workflow with draft upload, trigger detection, thumbnails, and per-request weights - Wildcard library with editable prompt tokens and encoder-assisted suggestion generation - Gallery indexing with favorites, dominant-color filters, bulk download, rebuild, and import support - Prompt enhancement and Creative Mode (`procedural_creativity` values `0-3`) - `R+` alternate staged inference mode with vibrance and bias controls - Content-aware x2 upscale with photo or illustration routing, tiled seam repair, and photo-only FS finishing - Source launchers for Windows, Linux, and macOS Tech stack: Python 3.11+ with FastAPI, Typer, PyTorch, Diffusers, Transformers, Accelerate, Pillow, SQLite, PowerShell, and Python helper scripts. Why it matters: JustRayzist simplifies the process of fine-tuning and generating images using Z-Image Turbo, making it accessible and efficient for developers and artists alike. Repository: MutantSparrow/JustRay

Overview: KVTC is an open-source implementation of NVIDIA's KVTC, designed to compress LLM KV caches by up to 6-9x. This reduces the memory bottleneck significantly, allowing for more concurrent users on the same hardware. Key features: - Compresses LLM KV caches by up to 6-9x with negligible quality loss. - Runs 2M+ token context on a single RTX 5090 GPU. - Integrates media-compression techniques inspired by JPEG and H.264. Tech stack: Python, PyTorch, CUDA. Why it matters: KVTC compresses LLM KV caches 6-9x, reducing memory usage and improving inference performance, especially for models like Qwen3.5-27B, which can handle 1M token context on a single RTX 5090 GPU. Repository: OnlyTerp/kvtc GitHub stars: 24 GitHub description: First open-source KVTC implementation (NVIDIA, ICLR 2026) -- 8-32x KV cache compression via PCA + adaptive quantization + entropy coding.

### Overview: This dataset contains over 700,000 Vietnamese vision-language samples, created by Gemini Pro. It includes several subsets for different tasks, such as conversation, complex reasoning, and detail description. The data is structured into five subsets and can be accessed through the Hugging Face dataset repository. ### Key Features: - **Caption-based Prompting**: Uses accurate captions and bounding boxes from the original dataset. - **Image-based Prompting**: Uses images to create captions and conversations. - **Translation for ShareGPT4V**: Employed translation prompts for the ShareGPT4V dataset. - **Curation Process**: Removed Han, Japanese, and Korean characters and refined samples with high perplexity levels. ### Tech Stack: Python ### Why It Matters: The dataset is crucial for researchers and developers looking to enhance their models with Vietnamese language capabilities, particularly in vision-language tasks. It provides a rich resource for training and evaluating models in Vietnamese, which is essential for applications in Vietnamese-speaking environments. ### Repository: The repository is hosted on GitHub at [Oztobuzz/Vista](https://github.com/Oztobuzz/Vista). ### GitHub Stars: The repository has been starred by 26 users, indicating its popularity and usefulness among the developer community. ### Dataset Structure: The dataset is organized into five subsets, each corresponding to a different task: - **Vi-LLAVA conversation**: 107,052 samples - **Vi-LLAVA complex reasoning**: 112,650 samples - **Vi-

### Overview: llmsearch is a Python tool designed for ML practitioners to optimize the generation parameters of large language models (LLMs). It simplifies the process of hyperparameter tuning by handling the setup and execution of grid or random search across various models and datasets. ### Key Features: - **Grid Search**: Automatically runs a grid search over multiple parameters to find the best set. - **Random Search**: Conducts random searches to explore a broader parameter space. - **Model Compatibility**: Supports a wide range of models and datasets. - **Performance Metrics**: Measures and tracks performance metrics such as accuracy, perplexity, etc. ### Tech Stack: - Python - PyTorch - Transformers - Pynvml for GPU stats ### Why it Matters: - **Efficiency**: Automatically handles the setup and execution of hyperparameter searches, saving time and effort. - **Scalability**: Supports a variety of models and datasets, making it versatile for different ML projects. - **Reproducibility**: Ensures consistent and reproducible results through detailed documentation and examples. ### Repository: - **Repository**: [Praful932/llmsearch](https://github.com/Praful932/llmsearch) - **GitHub Stars**: 26 - **Description**: Find better generation parameters for your LLM - **Documentation**: [Release blog](https://praful932.dev/blog-4-llmsearch/) ### Installation: - **Requirements**: `python>=3.8.1`, `torch>=

### Overview: FlowBP is a novel framework designed to improve the efficiency and effectiveness of reward backpropagation in text-to-image flow matching models. It separates the design space into four key choices: reward-model input, active timesteps, integration weights, and bridge coupling. This unified approach allows for a wide range of trade-offs in terms of accuracy, memory usage, and coupling, enabling strong cross-backbone gains across various models. ### Key features: - **Unified Design Space**: Separates direct reward backpropagation into four distinct choices. - **Three New Variants**: `FlowBP-Sparse`, `FlowBP-Bridge`, and `FlowBP-Lagrange` instantiate different accuracy, memory, and coupling trade-offs. - **Strong Cross-Backbone Gains**: Improves over the base model on all five HPDv2 metrics and leads most preference/quality scores. - **Baseline Coverage**: Includes direct-gradient baselines: `ReFL`, `DRaFT-LV`, `DRTune`, and `LeapAlign`. ### Tech Stack: Python ### Why it Matters: FlowBP provides a significant improvement in sample efficiency and effectiveness for reward backpropagation in text-to-image models, leading to strong cross-backbone gains across various models.

Overview: This project generates high-quality images with quotes, suitable for Instagram and Pinterest, in less than 5 seconds per 100+ images. Key features: - Generates images with quotes - Supports multiple topics (Christian, Fitness) - Uses Python for script execution - Uses PILLOW for text generation - Supports darkening images for better text visibility Tech stack: Python Why it matters: This script is useful for creating visually appealing images with quotes, making it perfect for social media platforms like Instagram and Pinterest. Repository: SamuraiPolix/Image-Quote-Generator GitHub stars: 26

### Overview: This project introduces **MemeCLIP**, a framework that leverages CLIP representations for multimodal meme classification. The framework is designed to address the challenge of understanding and classifying memes with complex textual and image content. It includes a novel dataset, PrideMM, which is specifically curated for the LGBTQ+ Pride movement, and a benchmarking experiment on existing datasets to evaluate its performance. ### Key Features: - **Datasets**: PrideMM, a dataset of text-embedded images associated with the LGBTQ+ Pride movement. - **Framework**: MemeCLIP, a novel framework that efficiently learns from pre-trained CLIP models to classify memes. - **Benchmarking**: Extensive experimentation on PrideMM and other datasets to establish benchmarks for each task. - **Performance**: Achieves superior performance compared to previously proposed frameworks on real-world datasets. ### Tech Stack: - **Python**: The primary programming language used for the project. ### Why It Matters: - **Enhanced Classification Accuracy**: MemeCLIP outperforms existing frameworks on real-world datasets, demonstrating its effectiveness in handling complex multimodal content. - **Novel Dataset**: PrideMM addresses a significant gap in existing resources, providing a comprehensive dataset for the study of LGBTQ+ related memes. - **Efficient Learning**: By leveraging pre-trained CLIP models, MemeCLIP can quickly learn from large-scale datasets, making it a powerful tool for meme classification tasks. ### Repository: - **GitHub Repository**: SiddhantBikram/MemeCLIP - **Stars**: 26 - **Description**:

Overview: DRIFT is a decentralized model that splits a single large language model across heterogeneous personal machines, including a Mac (Apple MPS) and a Windows PC (NVIDIA CUDA). It achieves this by layering the model and streaming only the hidden state between nodes over a framework-neutral byte protocol, ensuring identical results to running the model on a single machine. Key features: - Layered model splitting - Streamed hidden state between nodes - Framework-neutral byte protocol - Peer-to-peer hidden state - Encrypted and authenticated wire - Bitwise failover - Weightless head - Signed receipts Tech stack: Python, PyTorch MPS, CUDA, Apple MPS, NVIDIA CUDA Why it matters: DRIFT offers a unique solution for decentralized model deployment, eliminating the need for a datacenter and providing a framework-neutral way to run a single large language model across multiple machines. The peer-to-peer hidden state and encrypted wire ensure secure and efficient communication, while the framework-neutral byte protocol ensures identical results to running the model on a single machine. This approach enables decentralized model deployment without the limitations of traditional distributed systems.

Overview: This project introduces a loss function optimized for cross-entropy (CE) based on the Minimum Word Error Rate (MWER) Loss, incorporating beam search and negative sampling strategies. The project aims to improve the accuracy and efficiency of speech recognition tasks. Key features: - Smoothed Max Pooling Loss - MWER Loss with Beam Search and Negative Sampling Strategy Tech stack: Python Why it matters: This loss function offers a balance between computational efficiency and performance, making it suitable for real-time applications in speech recognition and related fields.

Overview: Proxima is an out-of-tree vLLM plugin that implements STAR-KV, a low-rank KV cache compression technique, to serve more concurrent requests and longer contexts in the same GPU memory. Key features: - Low-rank KV cache compression - Adaptive rank control via learned soft-thresholding - Per-layer low-rank projection for K, joint decomposition for V - Compressed paged format storage - Reproduces vLLM's pipeline with dedicated Triton kernels Tech stack: Python, vLLM, CUDA, Triton, ICML 2026 Spotlight Why it matters: Proxima allows existing GPUs to serve 4x more concurrent requests and longer contexts in the same GPU memory, improving performance and resource utilization.

### Overview: This repository contains the official code for our paper: "Thinking Before Looking: Improving Multimodal LLM Reasoning via Mitigating Visual Hallucination". The code is designed to enhance reasoning accuracy and reduce hallucinations in visual inputs by introducing the Visual Inference Chain (VIC) method, inspired by human cognition. ### Key Features: - **Visual Inference Chain (VIC)**: A novel approach to guide MLLMs to prioritize reasoning based on contextual knowledge before engaging with visual elements. - **GPT Demo**: An interactive GPT demo that allows users to explore the visual inference chain of a question. - **Method Description**: The paper outlines the method, including generating a Visual Inference Chain, extracting rationale instructions, and inferring the answer using the structured guidance provided by the Visual Inference Chain. ### Tech Stack: Python ### Why it Matters: This method aims to enhance the performance of MLLMs in complex tasks by reducing hallucinations and improving reasoning accuracy through a structured approach inspired by human cognition. ### GitHub Stars: 26 ### Repository: Terry-Xu-666/visual_inference_chain ### Primary Language: Python

### Overview: This repository contains the official dataset and model implementation for Natural Language Translation (NMT) for local languages in West Africa, specifically focusing on the Ewe language. The dataset is based on the EWE Corpus, and the Yodi model is developed using this dataset. ### Key Features: - EWE Corpus dataset for NMT - Yodi model implementation - Updated dictionaries for KABDICT525 and EWEDICT995 - Python modules for easy integration into projects ### Tech Stack: Python ### Why it Matters: This dataset and model are crucial for developing accurate machine translation and speech recognition capabilities for West African languages, particularly Ewe. The availability of labeled data and updated dictionaries ensures high-quality translation and recognition results. ### Repository: Umbaji/NMTMD GitHub Stars: 26 GitHub Description: Official repository for the Opensource Textdataset for NMT for local languages in West Africa (EWE Corpus) and implement the Yodi model afterward.

Overview: This repository contains the implementation for a paper presenting a regularized self-labeling approach aimed at improving the generalization and robustness of fine-tuned models in neural networks. Key features: - Layer-wise regularization - Self-labeling - Fine-tuning ResNet-101 on image classification tasks Tech stack: Python Why it matters: This method combines regularization and self-training techniques to enhance the performance and robustness of fine-tuned models, particularly in image classification tasks. Repository: VirtuosoResearch/Regularized-Self-Labeling GitHub stars: 26

### Overview: This repository contains curated research papers on Human-Computer Interaction (HCI), Large Language Models (LLM), Multi-Modal Language Models (MLLM), Agents, Retrieval-Augmented Generation (RAG), Agentic Reinforcement Learning (Agentic-RL), and Embodied AI (2021–present). The collection is organized into sections based on the papers' subjects and additional information provided. The repository includes quick start instructions for installing the CLI and searching for papers, as well as documentation for legacy scripts. ### Key Features: - **LLM**: Papers on LLMs, including error detection and assessment. - **MLLM**: Papers on Multi-Modal Language Models. - **Agent**: Papers on agents and their interactions. - **RAG**: Papers on Retrieval-Augmented Generation. - **Agentic-RL**: Papers on Agentic Reinforcement Learning. - **Embodied AI**: Papers on embodied AI. ### Tech Stack: Python ### Why It Matters: This collection provides a comprehensive overview of research in the field of human-computer interaction, large language models, and related technologies. It is useful for researchers, developers, and anyone interested in the latest advancements in these fields. ### GitHub Stars: 26 ### Repository: WILLOSCAR/Awesome-HCI-LLM ### Primary Language: Python ### GitHub Description: A curated collection of research papers on **HCI**, **LLM**, **MLLM**, **Agent**, **RAG**, **Agentic-RL**, and **Embodied AI** (20

### Overview: This project challenges developers to build an algorithm that can understand spoken commands using the TensorFlow framework. ### Key Features: - Train and test datasets available for Kaggle competition. - Comprehensive documentation and resources provided. - Detailed data description and evaluation metrics. - Prerequisite materials and tools for Kaggle competition. ### Tech Stack: Python ### Why it Matters: This project offers a practical application of TensorFlow in speech recognition, making it an excellent resource for developers interested in building voice recognition systems.

Overview: This project profile details the Qwen3.8-27B model optimized for 2× RTX 5090 GPUs, featuring a 262,144-token window, NVFP4 KV cache, and MTP head with 3 tokens per draft. It includes detailed measurements and benchmarks for various configurations, showcasing its effectiveness in serving concurrent coding tasks. Key features: - 262,144-token window - NVFP4 KV cache - MTP head with 3 tokens per draft Tech stack: Python Why it matters: This optimized model is designed for high-performance coding and prose generation, making it ideal for environments requiring fast and efficient processing of long context tasks. Repository: adrienbrault/qwen3.8-27b-rtx5090 GitHub stars: 26 GitHub description: Qwen3.8-27B on 2× RTX 5090 · 1.49M KV · code/prose t/s: c1 264/216, c16 2,649/2,240 · NVFP4 + vLLM + MTP

### Overview: This repository contains code for the findings of the ACL 2022 paper titled "Sentiment Word Aware Multimodal Refinement for Multimodal Sentiment Analysis with ASR Errors". The project aims to improve the accuracy of sentiment analysis in the presence of ASR errors. ### Key Features: - Sentiment Word Aware Multimodal Refinement - Multimodal Sentiment Analysis - ASR Error Mitigation ### Tech Stack: Python ### Why it Matters: This project addresses a significant challenge in natural language processing (NLP) by enhancing the accuracy of sentiment analysis in the presence of errors introduced by automatic speech recognition (ASR). The use of multimodal techniques and sentiment-aware refinement can lead to more reliable and accurate sentiment classification, which is crucial for applications such as customer feedback analysis, social media monitoring, and sentiment-based marketing strategies. Repository: albertwy/SWRM Primary language: Python GitHub stars: 26 GitHub description: Code for Findings of ACL 2022 Paper "Sentiment Word Aware Multimodal Refinement for Multimodal Sentiment Analysis with ASR Errors"

### Overview: OpenComputer is an open-source, local-first computer-use agent designed for Linux, Windows, and macOS. It leverages local Vision-Language Models (VLMs) or compatible APIs to provide a seamless, interactive experience. ### Key Features: - **Local by Default:** Screenshots, prompts, and tool results stay on your machine. - **Provider Independent:** Uses an OpenAI-compatible Chat Completions API. - **Cross-Platform Desktop Shell:** One Electron application for Linux, Windows, and macOS. - **Visible and Interruptible:** Watch actions, approve individual steps, or stop a running task. - **General Computer Control:** Works with normal desktop applications instead of browser-specific automation APIs. - **Open and Extensible:** Tools, model providers, safety policies, and integrations are regular TypeScript modules. ### Tech Stack: TypeScript ### Why it Matters: OpenComputer is experimental software with the ability to control your mouse and keyboard. Keep **Ask first** enabled, supervise actions, and do not use it around sensitive or irreversible workflows without understanding the risks. ### Repository: andykr1k/OpenComputer ### GitHub Stars: 26 ### Primary Language: TypeScript ### GitHub Description: Open-source local-first computer-use agent for Linux, Windows, and macOS, powered by local VLMs or compatible APIs.