Projects

### Overview: This repository tracks the progress in Entity Linking, focusing on recent trends and studies on Entity Representations. It includes contributions from various conferences and workshops, such as NAACL'21, ICLR'21, and others. The repository aims to provide insights into the latest advancements and challenges in the field of Entity Linking. ### Key Features: - **Trends (NAACL'21 and ICLR'21):** - **AUTOREGRESSIVE ENTITY RETRIEVAL:** Traditional entity linking systems face challenges like memory consumption, computational resource cost, and cold start problems. This method uses an autoregressive sequence-to-sequence approach to generate entity names. - **Linking Entities to Unseen Knowledge Bases with Arbitrary Schemas:** This approach converts the schema of unknown entities into BERT embeddings using attributes and auxiliary tokens, addressing the issue of unknown attributes. - **Trends (~EMNLP'20 and CoNLL'20):** - **In Media Res:** This corpus is designed to evaluate Named Entity Linking with creative works, providing annotations in different styles. - **LUKE:** This project proposes a pretraining task using BERT, incorporating an extended version of the transformer that considers entity-aware self-attention. - **Scalable Zero-shot Entity Linking with Dense Entity Retrieval:** This method addresses zero-shot entity linking by using dense entity retrieval. - **Entity Linking in 100 Languages:** This project introduces a new formulation for multilingual entity linking, resolving language

### Overview: This repository is a curated list of Legal Natural Language Processing (LegalNLP) resources, providing a wealth of datasets, tools, and resources for researchers and practitioners in the field of legal text analysis and prediction. ### Key Features: - **Datasets:** - **Legal Judgement Prediction (LJP):** Swiss court judgments (85K cases with 2 outcomes), EU court judgments (11K cases with 11 outcomes), Chinese court judgments (2.6M cases with 6 outcomes), and more. - **Legal Text Classification (LTC):** Greek legislation (47.5K laws with 2.7K labels), contracts (510 contracts with 41 classes), EU legislation (65K laws with 4.5K labels), contracts (60.5K contracts with 12.6K labels), contracts (2.6K clauses with 21 classes), EU legislation (57K laws with 4.3K labels), and more. - **Annotator Tools:** - AnnoCaseLaw: US Appeals Court negligence cases (471 annotated cases with expert labels). - IndianBailJudgments-1200: Indian court bail decisions (1.2K judgments with 20+ structured attributes). - CaseSumm: US Supreme Court opinions (25.6K opinions with official syllabuses). - JUSTICE: US Supreme Court cases (benchmark for judgment prediction). - Cambridge Law Corpus (

**Overview:** Monocle is a framework designed for tracing and testing GenAI applications written in Python. It provides rich traces, CI-friendly tests, and deep visibility across language models, agents, tools, and vector stores. **Key features:** - Monocle automatically instruments GenAI frameworks, creating spans for key operations without manual configuration. - It supports LangChain, LlamaIndex, Haystack, Google ADK, OpenAI, Anthropic, and more through first-class instrumentation. - Monocle emits standard OTLP spans, making it easy to send traces to observability backends. **Tech stack:** Monocle is built on OpenTelemetry, leveraging the Linux Foundation AI & Data community. **Why it matters:** Monocle helps developers ensure debuggability, reliability, and compliance for AI agents, making them easier to test and monitor. **Repository:** The repository is hosted on GitHub and has received 345 stars, indicating a strong community interest and usage. **Primary language:** The primary language used in Monocle is Python. **GitHub stars:** 345 **GitHub description:** Monocle is a framework for tracing GenAI app code. This repo contains implementation of Monocle for GenAI apps written in Python. **Repository:** monocle2ai/monocle **License:** Apache 2.0 **Runtime:** Python 3.7b **Frameworks:** Monocle supports LangChain, LlamaIndex, Haystack, Google ADK, OpenAI, Anthropic, and more. **

### Overview: HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination & Visual Illusion in Large Vision-Language Models ### Key Features: - HallusionBench is an image-context reasoning benchmark designed to challenge GPT-4V, LLaVA-1.5, and other multi-modality models. - Contains detailed analysis of examples demonstrating language hallucination and visual illusion in VLMs. - Provides insights into improving VLMs to mitigate these types of mistakes. - Challenges the state-of-the-art in image reasoning tasks. ### Tech Stack: Python ### Why It Matters: HallusionBench serves as an advanced diagnostic suite for evaluating and improving large vision-language models (VLMs) in handling entangled language hallucination and visual illusion. It provides a comprehensive benchmark that highlights the limitations of current models and offers practical insights for future research and model improvement.

Overview: Log Anomaly Detector is an open-source project designed to detect abnormal events in application logs using unsupervised machine learning. It includes components for training models, monitoring health, and a feedback system for correcting false predictions. Key features: - LAD-Core: Custom code for training and predicting anomalies using W2V and SOM models. - Metrics: Utilizes Grafana and Prometheus for monitoring the machine learning system's health. - Fact-Store: A metadata registry for tracking feedback and correcting false predictions. Tech stack: Python 3.6 Why it matters: This project aims to automate the detection of anomalies in application logs, helping developers quickly identify root causes or generate alerts for high-frequency anomalous logs, thereby improving application monitoring and root cause analysis. Repository: AICoE/log-anomaly-detector Primary language: Jupyter Notebook GitHub stars: 344

### Overview: DiffusionBERT is an improved generative masked language model based on discrete diffusion models, designed to enhance the quality of text generation. ### Key features: - DiffusionBERT combines the training strategies of diffusion models and pre-trained language models to improve text generation. - It explores the reverse process of a discrete diffusion model to train BERT, leading to significant improvements in perplexity and BLEU scores. - The model is trained with a new noise schedule for the forward diffusion process and various designs to incorporate time steps into BERT. ### Tech stack: Python ### Why it matters: DiffusionBERT offers a novel approach to generative language models by combining diffusion models and pre-trained language models, resulting in improved text generation quality and efficiency.

Overview: GPL is a powerful unsupervised domain adaptation method designed for training dense retrievers, specifically for dense retrieval tasks. It utilizes generative pseudo labeling to achieve significant improvements over zero-shot models. Key features: - GPL can be trained using only unlabeled data. - It leverages cross-encoders for effective domain adaptation. - The method requires minimal computational resources and is highly efficient. Tech stack: Python Why it matters: GPL is a significant advancement in unsupervised domain adaptation, offering a practical solution for improving dense retrieval models without the need for labeled data. This makes it particularly useful in scenarios where labeled data is scarce or expensive to obtain.

Overview: VIAME is a versatile computer vision application designed for DIY AI, offering features such as object detection, object tracking, data annotation, multi-camera processing, size measurement, image enhancement, rapid model generation, query-based search, and mosaicing. It includes tools for algorithm evaluation and supports both desktop and web interfaces. Key features: - Object Detection - Object Tracking - Data Annotation - Multi-Camera Processing - Size Measurement - Image Enhancement - Rapid Model Generation - Query-Based Search - Mosaicing - Pipeline Framework - Documentation Tech stack: Python Why it matters: VIAME is a comprehensive toolkit for computer vision tasks, making it useful for various applications and environments, including marine species analytics and generic computer vision applications. Repository: VIAME/VIAME GitHub stars: 343 GitHub description: Video and Image Analytics for Multiple Environments

Overview: This repository, GraphAware Natural Language Processing, provides Graph-Based Natural Language Processing capabilities for Neo4j, offering a unified interface and domain-specific language for developers. Key features: - Text information Extraction - Multi-languages in the same database - Custom NamedEntityRecognition model builder - ConceptNet5 Enricher - Microsoft Concept Enricher - Keyword Extraction - TextRank Summarization - Topics Extraction - Word Embeddings (Word2Vec) - Similarity Computation - PDF Parsing - Apache Spark Binding for Distributed Algorithms - Doc2Vec implementation - User Interface - ML Prediction capabilities - Entity Merging Tech stack: Java Why it matters: This plugin offers powerful NLP capabilities for Neo4j, making it easier and more efficient for developers to integrate natural language processing into their Neo4j applications. Repository: graphaware/neo4j-nlp GitHub stars: 344 GitHub description: NLP Capabilities in Neo4j

### Overview: Voost is a unified and scalable diffusion transformer designed for bidirectional virtual try-on and try-off, achieving high-quality results across various human poses, garment categories, backgrounds, lighting conditions, and image compositions. ### Key features: - **Unified Model**: Handles both virtual try-on and try-off within a single transformer. - **Scalability**: Efficiently scales to different resolutions and devices. - **Versatility**: Robust to various human poses, garment categories, backgrounds, lighting conditions, and image compositions. ### Tech stack: The project utilizes Hugging Face Spaces for a public demo and GitHub for version control and documentation. ### Why it matters: Voost is a significant advancement in virtual try-on technology, offering a unified approach to both try-on and try-off, making it more versatile and effective for a wide range of applications.

### Overview: MineBench is a Minecraft-style voxel benchmark designed to evaluate AI models' spatial reasoning abilities. It compares models' performance on natural-language prompts, generating precise 3D coordinates as JSON. The benchmark uses a global Bradley-Terry model to rank models based on their ability to construct complex voxel structures, providing insights into AI's spatial intelligence. ### Key Features: - **Arena Mode**: Models compete in blind head-to-head comparisons of pre-generated builds, with confidence-aware ranking. - **Sandbox Mode**: Allows direct comparison of existing builds, enabling users to visually assess model performance. - **Minimal Primitives**: Models can only use basic primitives like `block`, `box`, and `line` to generate large voxel structures. - **Global Bradley-Terry Model**: Ranks models based on their ability to construct voxel structures, providing a clear indication of their spatial reasoning skills. ### Tech Stack: - TypeScript ### Why It Matters: MineBench provides a unique way to assess AI models' spatial reasoning abilities, which correlates strongly with their general intelligence. It offers a practical way to visually determine the raw intelligence of models, making it useful for both academic and industrial use cases.

Overview: MegaDLMs is a GPU-optimized framework designed for training diffusion language models at any scale. It supports both dense and MoE architectures, offering comprehensive training pipelines and ultra-speed training speeds. Key features: - Full support for Diffusion Language Models (DLMs) and Autoregressive LMs - Up to 47% Model FLOP Utilization (MFU) and 3x faster training speed compared to other frameworks - Leverages Megatron-LM for flexible parallelism and GPU-optimized Transformer layers - Seamlessly integrates with HuggingFace checkpoints for seamless integration Tech stack: Python Why it matters: MegaDLMs is crucial for training DLMs and Autoregressive LMs at any scale, offering high performance and scalability, making it essential for researchers and developers working on large-scale language models. Repository: JinjieNi/MegaDLMs GitHub stars: 343 GitHub description: GPU-optimized framework for training diffusion language models at any scale. The backend of Quokka, Super Data Learners, and OpenMoE 2 training.

Overview: BMList is a comprehensive list of big models, featuring models with over 1 billion parameters that have been publicly released through papers, articles, or news. Key features: - Includes models with at least 1 billion parameters. - Offers a multi-dimensional table for easy classification by organizations, languages, domains, etc. - Features a gallery of figures showcasing statistics on big models. Tech stack: Python Why it matters: This repository provides a valuable resource for developers and researchers interested in big models, offering a centralized platform to explore and analyze these significant advancements in artificial intelligence.

### Overview: This repository is a collection of training data management resources for large language models, organized based on a survey paper. It includes curated papers on domain composition, data quantity, data quality, and relations among different aspects of training data management. ### Key Features: - **Domain Composition**: Papers on language models for dialog applications, data selection for language models via importance resampling, and lessons for training LLMs on programming and natural languages. - **Data Quantity**: Papers on scaling laws for neural language models, data repetition, scaling data-constrained language models, and document de-duplication and diversification. - **Data Quality**: Papers on quality filtering of text data and understanding data combinations for LLM training. ### Tech Stack: Python, PyTorch, Hugging Face Transformers, and various datasets. ### Why it Matters: This repository provides essential resources and insights for developers and researchers working on large language model training, helping them optimize their data management practices and improve the performance of their models.

### Overview: Darwinex Labs is a developer-focused project that provides datasets, tools, and resources related to Prop Investing and Quantitative Trading at Darwinex. The project aims to democratize the financial landscape by enabling traders and investors to access and trade in real-time, fair compensation for their work, and structured solutions for undercapitalized talent. ### Key Features: - **DWX_ZeroMQ_Connector**: A dedicated repository for future updates to the DWX_ZeroMQ_Connector project. - **Darwinex Labs**: Prop Investing Arm & Quant Team at Darwinex, providing access to datasets, tools, and more. - **DARWIN Asset**: A proprietary data wrapper that replicates alpha signal, enabling fair compensation and structured solutions for undercapitalized talent. - **Hedge Fund as a Service**: Enables traders to compete at institutional scale under Darwinex's regulatory umbrella. - **DARWIN Exchange**: An online marketplace for traders to list and trade their strategies, providing investors with a one-stop solution for Prime Brokerage, Risk Management, and Back-Office. - **Regulatory Compliance**: All transactions are protected by encryption and paid through a 20% success fee to the alpha generators. ### Tech Stack: - Python ### Why It Matters: Darwinex Labs is crucial for democratizing the financial landscape by enabling traders and investors to access and trade in real-time, while ensuring fair compensation and structured solutions for undercapitalized talent. The DARWIN Asset and DARWIN Exchange provide a platform for traders to list and trade their strategies, and investors to source talent

### Overview: The Ukrainian IT Communities repository is a comprehensive list of various Ukrainian IT-related communities, including news portals, Telegram groups, and more. It serves as a resource for individuals interested in staying updated on the latest developments in the Ukrainian IT industry. ### Key Features: - **IT in General (25)** - **Backend (31)** - **Databases (4)** - **Design (16)** - **DevOps (10)** - **Hardware & Embedded (10)** - **Frontend (23)** - **Game Development (18)** - **QA/QC (4)** - **Blockchain (5)** - **Industry News (9)** - **Artificial Intelligence (15)** - **Memes (2)** ### Tech Stack: HTML ### Why It Matters: This repository aims to provide a centralized platform for Ukrainian IT professionals and enthusiasts to connect, share knowledge, and stay informed about the latest trends and developments in the Ukrainian IT industry.

Overview: This repository contains the code for "ATISS: Autoregressive Transformers for Indoor Scene Synthesis," a paper presented at NeurIPS 2021. The code is designed to accompany the paper and includes detailed instructions for training models, using pretrained models, and performing interactive tasks. Key features: - Pretrained models for indoor scene synthesis - Interactive tasks for visualizing and manipulating indoor scenes - Code for training and evaluating models Tech stack: Python Why it matters: The code provides a practical implementation of the proposed method in the paper, allowing researchers and practitioners to easily reproduce and extend the results. Repository: nv-tlabs/ATISS GitHub stars: 343 GitHub description: Code for "ATISS: Autoregressive Transformers for Indoor Scene Synthesis", NeurIPS 2021

Overview: MeshXL is a generative pre-trained model family designed for 3D mesh generation, utilizing a Neural Coordinate Field representation. This model addresses the generation of unstructured 3D mesh data through modern LLM methods. Key features: - Neural Coordinate Field (NeurCF) representation - Explicit coordinate representation with implicit neural embeddings - Validation of NeurCF as a simple yet effective representation for large-scale sequential mesh modeling - Pre-trained weights available for different sizes (125M, 350M, 1.3B) on Hugging Face Tech stack: Python Why it matters: The model simplifies the generation of unstructured 3D mesh data, making it suitable for various applications in computer graphics, architecture, and more. Repository: OpenMeshLab/MeshXL Primary language: Python GitHub stars: 342 GitHub description: [NeurIPS 2024] MeshXL: Neural Coordinate Field for Generative 3D Foundation Models, a 3D fundamental model for mesh generation README excerpt: <!-- <img src="./assets/images/meshxl_logo.jpg" width="170px"> --> <div align= "center"> <h1> Official repo for MeshXL</h1> </div> <div align="center"> <h2> <a href="https://arxiv.org/abs/2405.20853">MeshXL: Neural Coordinate Field for Generative 3D Foundation Models</a></h2> <p

### Overview: RE2 is a fast and effective neural architecture designed for text matching tasks. It simplifies and optimizes the process by focusing on three key features: residual vectors, embedding vectors, and encoder outputs. RE2 achieves comparable performance to state-of-the-art models across various benchmarks, including SNLI, SciTail, Quora, and WikiQA, while significantly reducing inference time. ### Key features: - **Residual Vectors (R)**: Key features that capture the relationship between sequences. - **Embedding Vectors (E)**: Represents the input text sequences. - **Encoder Outputs (E)**: Provides contextual information for sequence alignment. ### Tech stack: Python ### Why it matters: RE2 simplifies and optimizes text matching tasks, making them faster and more efficient. It achieves strong performance on par with state-of-the-art models, with a notable 6x faster inference speed compared to other models.

Overview: Datmo is an open-source tool designed for data scientists to manage and track experiments in their projects. Key features: - **One command environment setup** (languages, frameworks, packages, etc) - **Tracking and logging** for model config and results - **Project versioning** (model state tracking) - **Experiment reproducibility** (re-run tasks) - **Visualize + export** experiment history - **(coming soon) Dashboards** to visualize experiments Tech stack: Python Why it matters: Datmo provides a streamlined way for data scientists to manage and track their experiments, ensuring reproducibility and project versioning, which is crucial for maintaining the integrity and traceability of their work.