A flaw was found in InstructLab. The `linux_train.py` script hardcodes `trust_remote_code=True` when loading models from HuggingFace. This allows a remote attacker to achieve arbitrary Python code execution by convincing a user to run `ilab train/download/generate` with a specially crafted malicious

A flaw was found in InstructLab. The `linux_train.py` script hardcodes `trust_remote_code=True` when loading models from HuggingFace. This allows a remote attacker to achieve arbitrary Python code execution by convincing a user to run `ilab train/download/generate` with a specially crafted malicious model from the HuggingFace Hub. This vulnerability can lead to …

CVE-2026-6855: A flaw was found in InstructLab. A local attacker could exploit a path traversal vulnerability in the chat session handler by manipulating the `logs_dir` parameter. This allows the attacker to create new directories and write files to arbitrary locations on the ...

Mark Kurtz explains how to overcome the technical and financial hurdles of scaling GenAI. He shares how to optimize LLM deployments with open-source tools, including vLLM for efficient serving, LLM Compressor for model compression, and InstructLab for fine-tuning with synthetic data. He provides a deep dive into balancing performance, accuracy, and cost to ensure successful production deployment. By Mark Kurtz

InstructLab.ai implements the large-scale alignment for the chatbots concept(LAB), which intends to overcome the scalability challenges in the instruction-tuning phase of a large language model…

The AI Innovation Team at Red Hat that builds InstructLab.

Plug your knowledge into a model easily with InstructLab MLPrague: 2025 Tomas Tomecek Senior Principal Software Engineer @TomasTomecek Karel Piwko Senior Principal Software Engineer @karel.piwko Cedri...

Most blog posts focus on using top-tier LLMs or setting up complex AI pipelines for large corporations. But what if your data is private, and you don’t have access to top-tier ML talent or massive infrastructure? In this article, we show how to fine-tune a model for mid-sized software development teams or IT support, using your own domain expertise. With **Apache Answer** and **InstructLab** , you can build a powerful, cost-effective AI solution tailored to your specific needs. ## InstructLab InstructLab is an open-source AI community project aimed at empowering individuals to shape the future of generative AI. It provides tools for users to fine-tune existing large language models (LLMs), such as Granite, using additional data sources. This allows LLMs to continuously gain new knowledge, filling in gaps from their initial training, including real-time updates on current events. Subject matter experts from any domain can contribute to enhancing the LLMs' knowledge. InstructLab’s collaborative platform fosters community-driven improvements and offers tools for experimenting with model updates and ensuring their quality. LAB: Large-Scale Alignment for ChatBots The **taxonomy YAML** in InstructLab is a structured file that contains a set of **question-answer pairs** , organized by domain, to represent specific **skills or knowledge**. Each YAML file includes metadata like version, task description, contributor info, and example prompts. InstructLab uses these YAML files to **generate synthetic training data** that fine-tunes **open-weight models** like Qwen or DeepSeek. By aligning the model with curated, domain-specific content, InstructLab helps improve accuracy and relevance in the model's responses across specialized topics. ## Apache Answer Apache Answer is an open-source AI project that enables users to fine-tune large language models (LLMs) with custom data. Its tools facilitate continuous model improvement, allowing users to fill gaps in existing knowledge and incorporate new information. Apache Answer offers a flexible platform for experimenting with model updates, ensuring improved response quality. Designed for accessibility and collaboration, the project encourages contributions to the advancement of generative AI. Apache Answer is similar to Stack Overflow in that it provides a platform for users to ask and answer questions, especially in technical or specialized domains. However, **unlike Stack Overflow, Apache Answer is open source and can be fully self-hosted on-premises** , giving organizations complete control over their data, customization, and user access. While Stack Overflow is a public, centralized platform mainly focused on general programming and tech topics, **Apache Answer can be tailored to any industry or organization**. It allows companies to build their own internal knowledge-sharing platforms—whether for software engineering teams, legal departments, medical institutions, or customer support operations. In essence, Apache Answer offers the same collaborative Q&A experience as Stack Overflow but with **full ownership, flexibility, and adaptability** for private or specialized use. ## Synthetic data Synthetic data generation is crucial for industries where real-world data cannot be used due to privacy or regulatory concerns. **Apache Answer** , a Q&A platform similar to StackOverflow, allows organizations to gather relevant industry-specific data and insights from codebases or domain experts. This data can be leveraged for generating high-quality synthetic datasets tailored to specific business needs. **InstructLab** enhances this process by fine-tuning models to produce contextually accurate synthetic data, ensuring it reflects real-world scenarios. Combined with **RAG** (Retrieval-Augmented Generation) and **CAG** (Cache-Augmented Generation), **InstructLab** enables both real-time and cached synthetic data generation. YAML files produced by **Apache Answer** contain structured question-answer data that can be used to **fine-tune open-source language models**. By feeding this data into training workflows, models like **Qwen** or **DeepSeek** can be aligned with specific domains or organizational knowledge, improving their accuracy and relevance. Example YAML Q&A ## RAG & CAG **RAG** (Retrieval-Augmented Generation) is an AI method that finds the most up-to-date information from external sources every time a question is asked. It combines a search system with a language model so the answers are both current and accurate. **CAG** (Cache-Augmented Generation) is different. It collects and stores all the needed information in advance. The model then uses this stored data to give fast answers without needing to search each time. With **RAG** , InstructLab helps improve the model’s ability to follow instructions _after_ retrieving fresh information. Since RAG fetches real-time data, the model needs to blend that information into a clear, helpful response. InstructLab’s fine-tuning methods make the model better at using that external content effectively. With **CAG** , where information is pre-loaded into a cache, InstructLab helps train the model to give accurate and helpful answers _from that fixed data_. It can improve how well the model uses the cached knowledge by fine-tuning it on examples that mirror expected questions and answers, ensuring faster and more relevant results. In short, InstructLab makes models better at following instructions and delivering useful answers — whether the data comes in real-time (RAG) or from a pre-built cache (CAG). ## Deployment InstructLab can be installed locally using `pip` or the `uv` package manager, making it easy to set up for individual use or testing. Alternatively, it can run inside a Docker container, providing a consistent environment for development and deployment. For scalable production environments, InstructLab can also be deployed on Kubernetes, leveraging its orchestration capabilities to handle scaling and resource management efficiently. This flexibility ensures it adapts to various workflows, from local experimentation to large-scale distributed deployments. To deploy InstructLab on Kubernetes, use a Makefile to define and manage Kubernetes resources like Deployments and ConfigMaps as multi-line variables. Dynamically generate labels and annotations using Git metadata and environment variables. Validate configurations, enforce constraints, and apply manifests with `kubectl`. This approach avoids Helm's complexity while maintaining flexibility and transparency. Time to replace Helm: Back to the Future git clone https://github.com/avkcode/InstructLab.git Using kubectl: kubectl apply -f instructlab.yaml Using make: make => InstructLab Management System Available targets: Deployment: deploy - Deploy InstructLab with ConfigMap undeploy - Remove InstructLab deployment Interaction: logs - View container logs status - Show deployment status Utility: help - Show this help message With tools like **Apache Answer** and **InstructLab** , you can fine-tune models using your domain knowledge without needing vast resources. You don't need to be a large corporation or have a huge ML team to harness the power of AI for your specific needs.

Red Hat is hiring for Full Time Software Quality Engineer - Boston, United States, a mid-level AI/ML/Data Science role offering benefits such as 401(k) matching, career development, equity / stock options, flex hours, flexible spending account, flex vacation, health care, medical leave, parental leave, salary bonus, transparency.

Red Hat Summit is the premier enterprise open source event for IT professionals

announce

🔰概要 InstructLabは、大規模言語モデル（LLM）に新しい知識やスキルを学習させるためのフレームワークです。 GPT-4などの高価なモデルを使わずに、独自のタクソノミー構造と合成データ生成により、安価かつ柔軟にモデルを微調整できるのが特徴です。 本記事では、InstructLabを実際に動かすためのDocker環境構築手順を紹介します。 🎯 目的（執筆の背景） Instruc...