Project directory

Projects

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

Python 7 stars

Phoenix-chaser/CodeAgent-R

Auditable code-repair post-training platform: source-group-isolated QLoRA SFT, mutation-audited tasks, execution-based checkpoint selection, paired frozen-test evaluation. 3-seed protocol on Qwen2.5-Coder-7B: mean +1.7pp, zero regressions

Overview: CodeAgent-R is an auditable code-repair platform that leverages QLoRA SFT and mutation-audited tasks to isolate source-group isolation. It includes a 3-seed protocol for Qwen2.5-Coder-7B, achieving a mean +1.7pp improvement and zero regressions. Key features: - Source-group isolation - Mutation-audited tasks - Execution-based checkpoint selection - Paired frozen-test evaluation Tech stack: Python Why it matters: This platform provides a reliable and auditable way to improve code quality through isolated training and mutation-audited tasks, ensuring that the modifications made are effective and do not regress. Repository: Phoenix-chaser/CodeAgent-R Primary language: Python GitHub stars: 5 GitHub description: Auditable code-repair post-training platform: source-group-isolated QLoRA SFT, mutation-audited tasks, execution-based checkpoint selection, paired frozen-test evaluation. 3-seed protocol on Qwen2.5-Coder-7B: mean +1.7pp, zero regressions.

TypeScript 7 stars

Prometheus-000/visionary-platform

Single-user training and generation platform on Modal. Train Krea 2 LoRAs, generate stills, and turn them into video with MiniMax-H3 (sound and picture in one pass) — one deploy, one URL, no infrastructure to keep alive.

### Overview: A generative studio that leverages Modal for single-user training and generation, allowing users to train Krea 2 LoRAs, generate stills, and create video clips using MiniMax-H3. The platform is designed to be deployed with a single URL, eliminating the need for maintaining infrastructure. ### Key Features: - **Training and Generation:** Train Krea 2 LoRAs on user-provided photographs and generate video clips using MiniMax-H3. - **Single-User Deployment:** Deploy the platform with a single URL, eliminating the need for ongoing maintenance. - **No Infrastructure:** The platform runs on Modal without requiring any additional infrastructure setup. ### Tech Stack: TypeScript ### Why It Matters: This platform simplifies the process of creating video content by combining photography, video editing, and AI, all accessible through a single URL. It eliminates the need for complex infrastructure, making it ideal for developers and creatives looking to streamline their workflow.

Python 7 stars

PromptPotter/prompt-potter-optimizer

LLM-driven evolution of prompts and pipelines. Evolves the prompt and the pipeline's parameters together against your labeled data, cuts losing candidates early, and shows every measurement behind the winner. Open source.

### Overview: PromptPotter is an LLM-driven optimization tool designed to enhance the performance of prompts and pipeline parameters using labeled data. It evaluates candidates early to ensure the best results and provides detailed measurements for the chosen configuration. ### Key Features: - **Prompt Optimization**: Automatically evolves the prompt and pipeline parameters together, cutting losing candidates early. - **Pipeline Parameter Tuning**: Refines the pipeline's parameters based on the evolving prompt. - **Detailed Measurements**: Shows every measurement behind the winning configuration. ### Tech Stack: - Python ### Why It Matters: PromptPotter leverages LLMs to optimize prompts and pipeline parameters, ensuring the best performance and efficiency. It cuts down on unnecessary iterations and provides clear, actionable insights, making it a valuable tool for developers and data scientists. ### Repository: - **Repository**: PromptPotter/prompt-potter-optimizer - **Stars**: 7 - **Description**: LLM-driven evolution of prompts and pipelines. Evolves the prompt and the pipeline's parameters together against your labeled data, cuts losing candidates early, and shows every measurement behind the winner. Open source. ### Works with: - **Agent Frameworks**: LangGraph, CrewAI, OpenAI Agents SDK, Bedrock Agents, n8n, Dify, FastAPI. - **DSPy**: `PromptPotterOpt`, GEPA. - **DSPy**: Each candidate runs as a Harbor trial, graded by the task's own checker. - **DSPy**: Langfuse / MLflow: Every round fans out as traces; the winner lands in ML

Python 7 stars

Redhatvale/4xV100-qwen38-flash-next-abliterated-128gb-vram

Runbook + benchmarks: Qwen3.8-Flash-Next-ABLITERATED NVFP4 on 4× Tesla V100-32GB (reflashed SXM2→PCIe, 2+2 NVLink + PLX). 1Cat-vLLM 1.5.0, TP4 — 262,144-token context validated, 46 tok/s decode, 122 tok/s aggregate at 4 concurrent streams. Full E0–E17 optimization log with measured evidence.

Overview: This repository documents the build process and performance metrics of a 4× Tesla V100-32GB GPU setup running the Qwen3.8-Flash-Next-ABLITERATED model, optimized for NVFP4 architecture. The model achieves a context size of 262,144 tokens with a decoding speed of 57 tokens per second, making it highly efficient and ready for use. Key features: - 4× Tesla V100-32GB GPUs - 2+2 NVLink + PLX interconnect topology - Full optimization program with 18 stages - 262,144-token context size validated - 46 tokens per second decode speed - 122 tokens per second aggregate decode speed at 4 concurrent streams Tech stack: NVIDIA ModelOpt, 1Cat-vLLM 1.5.0, NVFP4 weights Why it matters: This repository provides a comprehensive guide to the build process, including every step and fix, ensuring that the model can be deployed efficiently and effectively.

TypeScript 7 stars

Resk-Security/resk-llm-ts

A comprehensive security toolkit for JavaScript applications using Large Language Models. This library provides a wrapper around OpenAI-compatible APIs with built-in security features to protect your LLM integrations.

Overview: Resk-LLM-TS is a comprehensive security toolkit designed specifically for JavaScript applications using Large Language Models (LLMs). This library provides a wrapper around OpenAI-compatible APIs with built-in security features to protect your LLM integrations. Key features: - Zero dependencies - One pipeline for prompt injection defense - Built-in security features - Zero runtime dependencies - TypeScript support Tech stack: TypeScript Why it matters: Resk-LLM-TS offers a robust solution for security-conscious developers who want to integrate LLMs into their JavaScript applications while ensuring the integrity and security of their models. It simplifies the process of protecting against potential threats and provides a clean, secure interface for developers to work with. Repository: Resk-Security/resk-llm-ts GitHub stars: 7 GitHub description: A comprehensive security toolkit for JavaScript applications using Large Language Models. This library provides a wrapper around OpenAI-compatible APIs with built-in security features to protect your LLM integrations.

Python 7 stars

RobertoDeLaCamara/Research-Agent

Autonomous research agent — 10 sources in parallel, self-correcting loop, exports PDF/Word/HTML. Powered by LangGraph + Ollama.

Overview: Research-Agent is an autonomous research agent that searches 10 sources in parallel, fact-checks the synthesis, and exports reports in PDF, Word, HTML, or Markdown formats. It uses LangGraph and Ollama for its functionality. Key features: - **Parallel Multi-Source Research**: Searches 10 sources concurrently. - **Research Personas**: Supports 7 different research personas. - **Local Knowledge (RAG)**: Uploads PDFs or TXT files for indexed retrieval. - **Fact-Check Layer**: Evaluates for dubious claims and flags them. - **Cross-Session Memory**: Retains past research for future use. - **Report Metadata**: Includes word count and reading time. - **Export Center**: One-click export to multiple formats. - **MCP Server**: Exposes the agent via Model Context Protocol. - **Configurable Depth**: Quick, Standard, or Deep. - **Multilingual**: Supports English for global coverage. - **UI Language Switcher**: Switches between English and Spanish. - **Cloud LLM Support**: Compatible with various OpenAI-compatible APIs. Tech stack: LangGraph + Ollama. Why it matters: Research-Agent streamlines research and fact-checking processes, making it useful for various roles like market analysts, software architects, and news editors.

Python 7 stars

SSHdotCodes/deepseek-v4.1-flash-pro6000

DeepSeek-V4.1-Flash-1017 at native precision on one RTX PRO 6000 + 128 GB RAM: 31.6 tok/s decode, 1M context. ds41d engine with VRAM/RAM/NVMe expert tiers, GPUDirect Storage and DSpark speculative decoding

Overview: DeepSeek-V4.1-Flash-1017 is a highly precise model optimized for one RTX PRO 6000 GPU, achieving 31.6 tok/s decode performance. It runs on 128 GB of RAM and uses GPUDirect Storage and DSpark speculative decoding. Key features: - Native precision decoding - 1,048,576-token context - Low, High, and Max reasoning effort - Tool calls - Vision tower from the checkpoint - Speculative decoding - OpenAI Chat Completions, Anthropic Messages, and OpenAI Responses APIs - 3-bit experts for faster decoding Tech stack: Python Why it matters: This model is designed for high-performance decoding on a single RTX PRO 6000 GPU, providing exceptional speed and precision, making it ideal for real-time applications and high-throughput tasks. Repository: SSHdotCodes/deepseek-v4.1-flash-pro6000 GitHub stars: 6 GitHub description: DeepSeek-V4.1-Flash-1017 at native precision on one RTX PRO 6000 + 128 GB RAM: 31.6 tok/s decode, 1M context. ds41d engine with VRAM/RAM/NVMe expert tiers, GPUDirect Storage and DSpark speculative decoding

Java 7 stars

SUONSUN9527/AImv

muti-agent-仿seedance图片视频生成平台

Overview: AImv is a provider-agnostic multi-agent creative generation platform that leverages RAG (Revised Retrieval-Augmented Generation) and structured LLM output to generate images or videos based on a user-provided creative prompt. It includes a fixed pipeline with 7 stages, quality review, and continuous multi-round dialogues with context compression for improved quality and reliability. Key features: - Multi-stage Agent Chain: 7 stages for generating images or videos. - Real RAG: pgvector vector retrieval and keyword mixing with rerank. - Structured LLM Output: JSON schema with error handling and retry. - Provider-agnostic: Supports three gateways: DashScope, Pollinations, and Fixture. - Quality Gate: Auto-retry for non-compliant outputs and switch to backup credentials for quality review. - Continuous Multi-Round Dialogues: Context compression for improved quality and reliability. Tech stack: Java 17, Spring Boot, Vue 3, PostgreSQL 16, pgvector. Why it matters: AImv provides a robust, provider-agnostic, and quality-controlled workflow for generating images and videos based on user input, ensuring consistent and reliable results through a fixed pipeline and quality gate. Repository: SUONSUN9527/AImv GitHub stars: 7 GitHub description: A multi-agent creative generation platform.

Python 7 stars

Saganaki22/ComfyUI-Seedream4_Replicate

A ComfyUI custom node for ByteDance's Seedream 4 image generation model via the Replicate API. Generate high-quality images with advanced control over aspect ratios, multi-image generation, and seed management.

### Overview: This ComfyUI custom node allows developers to generate high-quality images using ByteDance's Seedream 4 image generation model via the Replicate API. It offers advanced control over aspect ratios, multi-image generation, and seed management, making it ideal for creating detailed and varied image sets. ### Key Features: - **Multiple Aspect Ratios**: Preset sizes for common aspect ratios. - **Sequential Image Generation**: Supports related image sets with prompts like "story sequence" or "character variations". - **Image-to-Image Support**: Works with reference images for image-to-image generation. - **Seed Control**: Allows reproducible results through seed management. - **Real-Time Progress Tracking**: 4-stage progress updates for easy monitoring. - **Automatic Cancellation**: Support for canceling image generation with ComfyUI's cancel button. - **Error Handling**: Detects and handles content policy issues. ### Tech Stack: Python ### Why It Matters: This node provides a powerful tool for generating high-quality images with advanced control, making it useful for various creative and design projects.

Python 7 stars

Sakura66/sagesched

SageSched: Intelligent LLM Request Scheduler with Workload Prediction — QoS-aware dual-queue scheduling for black-box LLM APIs (OpenAI/Azure/Doubao/Gemini)

Overview: SageSched is an intelligent scheduling system designed for black-box LLM APIs, predicting request workload before execution to ensure QoS guarantees. Key features: - Workload Prediction - QoS-Aware Scheduling - Adaptive Congestion Control - Multi-Account Load Balancing - Gittins Index Scheduling - Zero-GPU Required Tech stack: Python Why it matters: SageSched addresses common challenges in calling LLM APIs, such as blocking short requests, predicting costs, managing congestion across multiple accounts, and providing real-time visibility into request behavior. Repository: Sakura66/sagesched GitHub stars: 7

Jupyter Notebook 7 stars

SayamAlt/Language-Detection-using-fine-tuned-XLM-Roberta-Base-Transformer-Model

Successfully developed a language detection transformer model that can accurately recognize the language in which any given text is written.

Overview: Successfully developed a language detection transformer model that can accurately recognize the language in which any given text is written. Key features: - Accurate language detection - Fine-tuned XLM-Roberta-Base Transformer Model - Easy to use in Jupyter Notebook Tech stack: Jupyter Notebook Why it matters: This model can be used in various applications requiring language identification, such as translation tools, content localization, and more.

Python 7 stars

Scottcjn/legend-of-elya-genesis

Edge MoE inference on 1988 silicon: an integer-only ternary transformer live on a Sega Genesis 68000 — memory-mapped cartridge ROM makes expert activation free. Zero-cost expert routing, measured.

### Overview: ELYA INTO DREAMS is a pioneering project that demonstrates edge model parallelism (MoE) inference on a Sega Genesis 68000, achieving impressive results with a simple, integer-only ternary language model. The project showcases the effectiveness of zero-cost expert routing, measured, and the model's ability to generate fluent, related, and correct responses without precomputed text or lookup tables. ### Key Features: - **Edge MoE Inference**: ELYA INTO DREAMS leverages edge model parallelism (MoE) to achieve live inference on a 68000 processor. - **Zero-Cost Expert Routing**: The model uses zero-cost expert routing, measured, to activate experts without copying or using up precious RAM. - **Memory-Mapped ROM**: The model runs directly from the cartridge ROM, which is memory-mapped, allowing it to avoid the pitfalls of traditional CPU-based models. - **Fluent and Correct Responses**: The model generates fluent, related, and correct responses, even when it thinks too quickly or too slowly. - **Real Hardware Demonstration**: The project features a full video of ELYA INTO DREAMS running on a real Sega Genesis, showcasing its capabilities. ### Tech Stack: Python ### Why It Matters: The project challenges the conventional wisdom that activation cost is solely determined by the model architecture, proving that it is a property of the address model. By demonstrating that zero-cost expert routing is possible on a 68000, ELYA INTO DREAMS provides a valuable insight

Jupyter Notebook 7 stars

Spectrewolf8/PHi-3-SQL-generation-fine-tune-experiment

A fine-tuned version of Phi-3-mini-4k-instruct for generating SQL queries from natural language prompts, utilizing synthetic datasets and QLoRA for efficient adaptation and deployment.

### Overview: This repository contains a fine-tuned version of Phi-3 Mini, a smaller variant of the Phi-3 series, designed to generate SQL queries from natural language prompts. The model is optimized for efficient adaptation and deployment, utilizing synthetic datasets and QLoRA techniques. ### Key features: - SQL query generation - Natural language instruction translation - Fine-tuned for efficient adaptation and deployment ### Tech stack: - Jupyter Notebook ### Why it matters: This model is particularly useful for applications requiring interactive and responsive dialogue systems, such as chatbots or virtual assistants that interact with databases. It can also be integrated into tools designed for automatic query generation based on user-friendly descriptions. ### GitHub stars: 7 ### Repository: Spectrewolf8/PHi-3-SQL-generation-fine-tune-experiment

HTML 7 stars

SuperInstance/SuperInstance

Overview: SuperInstance is a system designed to integrate intelligence into existing work environments, such as boats, workshops, and households, by using small, verified parts that compose like spreadsheet formulas, perceive continuously, and can show receipts. The system is built from open-source development, measured against reality, and extracted into standalone tools when they prove themselves. Key features: - Intelligence integration into existing work environments - Continuous perception of environments - Receipts showing the value of the system - Open-source development and testing - Reproducibility and portability of the system - Composability of the system Tech stack: Quilt Why it matters: SuperInstance leverages the proven success of the spreadsheet as a programming interface, extending it to cells that can be anything from a sensor to an API call, ensuring that the system is reliable, reproducible, and portable. The use of Quilt as the underlying technology ensures that the system can be written once, verified once, and run anywhere, making it highly scalable and adaptable to different environments. Repository: SuperInstance/SuperInstance Primary language: HTML GitHub stars: 7 GitHub description: No description provided. README excerpt: # SuperInstance ![F/V Eileen — relational intelligence runtime, cutaway and flow-state cross-section: the 10-year invariant, reality is a database in the engine room](top.jpg) [a note from the picture-maker →](https://github.com/SuperInstance/AI-Writings/blob/master/first-take/metas-first-take.md) **The spreadsheet that thinks.

Python 7 stars

SynapCores/synapcores-agent

Real, framework-free AI support agent where SynapCores is the brain — memory, RAG, tool routing, generation in one database. Browser chat widget + live Brain debug sidebar. Fork and run in 30s.

Overview: SynapCores is a framework-free AI support agent that integrates memory, RAG, tool routing, and generation within a single database. This agent is designed to be embedded on any website using a single `<script>` tag, providing a browser chat widget with a live Brain debug sidebar. Key features: - Real, memory-grounded AI support - MIT-licensed and CDN-hosted chat UI - Native agent loop running via `AGENT_RUN()` in SynapCores - Lightweight proxy to handle SynapCores credentials and WebSocket communication - Supports bare HTML, Next.js, and WordPress embeds Tech stack: Node.js and TypeScript Why it matters: This agent simplifies the integration of AI support into websites, making it easy to embed and manage without the complexity of traditional AI frameworks. The lightweight architecture ensures minimal impact on the browser and database, making it suitable for a wide range of applications. Repository: SynapCores/synapcores-agent GitHub stars: 7 GitHub description: Real, framework-free AI support agent where SynapCores is the brain — memory, RAG, tool routing, generation in one database. Browser chat widget + live Brain debug sidebar. Fork and run in 30s.

Python 7 stars

TammineniTanay/hybrid-rag-system

Production-grade RAG system with hybrid retrieval (Qdrant + Elasticsearch + Neo4j), Corrective RAG via LangGraph, feedback-driven reward model, and RAGAS evaluation dashboard

Overview: This project is a production-grade Retrieval Augmented Generation (RAG) system that leverages hybrid retrieval strategies, including dense embeddings, BM25 sparse search, and knowledge graph traversal. It implements Corrective RAG (CRAG) for self-healing retrieval and continuous learning through user feedback, providing a real-time evaluation dashboard for faithfulness, relevance, and precision metrics. Key features: - Corrective RAG (CRAG) for self-healing retrieval - Hybrid retrieval strategies (dense embeddings, BM25, Neo4j) - Continuous learning through feedback - Production-grade system with a real-time evaluation dashboard Tech stack: Python Why it matters: This system addresses the challenges faced by production teams at large language model (LLM) companies, such as Anthropic and Cohere, by providing a robust and adaptable RAG system that can autonomously correct retrieval failures and continuously improve based on user feedback. Repository: TammineniTanay/hybrid-rag-system GitHub stars: 7 GitHub description: Production-grade RAG system with hybrid retrieval (Qdrant + Elasticsearch + Neo4j), Corrective RAG via LangGraph, feedback-driven reward model, and RAGAS evaluation dashboard. The system is part of the UniLLMOps framework and includes 50-Query Controlled Validation Results.

Python 7 stars

TiniLLM/ollama-token-monitor

Real-time terminal dashboard for Ollama — live output token counting (tokens/sec), CPU & thread usage, Apple Silicon GPU utilization, VRAM and loaded models. htop for your local LLM.

### Overview: Real-time terminal dashboard for Ollama — live output token counting, CPU & thread usage, GPU & VRAM monitoring. ### Key features: - **Live Token Tracking**: Output tokens and prompt tokens counted as they stream, per-request tok/s, session totals, and a tok/s sparkline. - **Per-Request Table**: Model, endpoint, output tokens, tokens/sec, elapsed time, and state (`generating`, `thinking/loading`, `done`). - **CPU & Threads**: Total and per-core CPU, load average, plus CPU%, thread count, and memory of every Ollama process. - **GPU Monitoring on Apple Silicon**: Device/renderer/tiler utilization % and GPU memory in use, read from `ioreg` (no sudo required). - **Loaded Models**: Size, VRAM, GPU/CPU placement split, context length, and keep-alive expiry countdown (from `/api/ps`). - **Single File, No Daemon**: One Python script using `rich`, `psutil`, `aiohttp`. ### Tech stack: Python ### Why it matters: Real-time monitoring and analysis of Ollama's performance metrics, including token counts, CPU and thread usage, GPU and VRAM utilization, and loaded models.

Python 7 stars

TruongNV-hut/AIcandy_LLM_Finetuning_bloom_560m_iehimqko

Fine-tuning large language models

Overview: Fine-tuning large language models involves adapting pre-trained models to specific tasks or domains by training them on smaller datasets, enhancing their performance for targeted applications. Key features: - Fine-tuning large language models (LLMs) for specific tasks - Leveraging pre-trained models' general knowledge for specialized applications - Techniques like supervised fine-tuning, reinforcement learning, and prompt engineering - Improving performance in sentiment analysis, text summarization, and domain-specific applications Tech stack: Python Why it matters: Fine-tuning LLMs is a cost-effective method to maximize the model's utility for targeted applications, making it a valuable technique in AI and ML. Repository: TruongNV-hut/AIcandy_LLM_Finetuning_bloom_560m_iehimqko Primary language: Python GitHub stars: 7 GitHub description: Fine-tuning large language models README excerpt: Fine-tuning large language models involves adapting pre-trained models to specific tasks or domains by training them on smaller datasets, enhancing their performance for targeted applications.

Unknown 7 stars

Vill-Lab/2024-TPAMI-VGDCU

Benchmark for VG-based Detection and Chart Understanding (IEEE T-PAMI 2024)

### Overview: This repository contains a benchmark dataset for the detection and chart understanding of vector graphics (VG-DCU), focusing on vector-based detection and chart understanding. The dataset consists of over 10,000 vector charts, including synthetic and real-world subsets, and is designed to advance the field of vector image detection. ### Key Features: - **VG-DCU Dataset**: A large-scale benchmark dataset for vector graphics detection and chart understanding. - **Synthetic and Real-World Subsets**: Contains 10,682 synthetic charts and 4,515 real charts, with 80% for training and 20% for testing. - **Chart Type and Distribution**: Includes charts of various types such as Area, Bar, Box, Donut/Pie, Line, Scatter, Heatmap, Counter, Violin, and Sankey, with detailed statistics on categories and bounding box instances. - **ChartQA and PlotQA**: Contains datasets for Chart QA and Plot QA tasks, focusing on the detection of charts and plots in vector graphics. ### Tech Stack: - **Vector Graphics**: Utilizes vector graphics for charts and plots. ### Why it Matters: The benchmark dataset is crucial for researchers and developers in the field of computer vision and image processing, as it provides a comprehensive and diverse dataset for evaluating and advancing vector-based detection and chart understanding.

Python 7 stars

VirtuosoResearch/Robust-Fine-Tuning

Measuring generalization properties of fine-tuning using Hessian

Overview: This repository provides an implementation for evaluating Hessian-based measurements of fine-tuned models, focusing on the generalization properties of fine-tuned models using Hessian distances. The results are based on the work presented at ICML'22 and DC, emphasizing the nonvacuous measure of generalization gaps. Key features: - Fine-tuning ResNet models under label noise - Fine-tuning Vision Transformer on noisy labels - Evaluating Hessian-based measures for noise stability and loss's Hessian trace Tech stack: Python Why it matters: By incorporating Hessians with distance from initialization, the Hessian distance measure provides a nonvacuous measure of the generalization gaps of fine-tuned models, highlighting the importance of these measures in understanding the generalization capabilities of fine-tuned models.