Releases·Tool Release
NVIDIA open-sources NeMo Data Designer, a framework for multimodal synthetic data generation
A paper on arXiv (2026-09-15) presents NeMo Data Designer (NDD), an open-source framework for multimodal synthetic data generation. Users declaratively define dataset columns (text, code, structured outputs, images, embeddings, statistical samplers), with a plugin system and a built-in preview-and-revision loop. It matters to practitioners because it was used to produce datasets for Nemotron model development and production enterprise deployments.
0 engagement1 sourceTue, Sep 15, 2026, 06:11 PM
AI Overview
- NeMo Data Designer is an open-source framework generating synthetic data via declarative column configs spanning text, code, images, embeddings, and samplers
- NDD configs are inspectable artifacts enabling reproducibility, with a plugin system for custom column types
- NDD builds in a preview-and-revision loop and handles dependency resolution, model endpoint calls, and request retries at runtime
- Case studies include datasets used in Nemotron model development and production enterprise deployments
Related
Releases·Model Release
Salesforce and Nvidia’s new reasoning model is everything the AI labs should fear
Salesforce announced Koa, its first reasoning model, at Dreamforce, co-developed with Nvidia on Nvidia's open-weight Nemotron base and post-trained for sales, marketing, and customer-support tasks. Koa will serve as an in-house alternative to frontier models like Claude and ChatGPT within Salesforce's Agentforce platform, trained on synthetic data with no customer data ingested. Nvidia claims a token-efficient inference architecture, positioning Koa as cheaper and more sovereign than routing enterprise agent tasks to frontier labs.
Salesforce announced Koa, its first reasoning model, at Dreamforce, co-developed with Nvidia on Nvidia's open-weight Nemotron base and post-trained for sales, marketing, and customer-support tasks. Koa will serve as an in-house alternative to frontier models like Claude and ChatGPT within Salesforce's Agentforce platform, trained on synthetic data with no customer data ingested. Nvidia claims a token-efficient inference architecture, positioning Koa as cheaper and more sovereign than routing enterprise agent tasks to frontier labs.
0 engagement2 sourcesrssTue, Sep 15, 2026, 12:00 PM
Developments·Paper
Researchers adapt NVIDIA Nemotron retrieval stack for Modern Greek, release HERA benchmark
A research team adapted NVIDIA's Nemotron retrieval stack for Modern Greek, covering corpus mining, synthetic supervision, retrieval training, reranker adaptation, and reader fine-tuning, and introduced a new benchmark called HERA. The work addresses the absence of Modern Greek from Nemotron retrieval models and major multilingual benchmarks, targeting RAG in legal, energy, financial, and medical domains. Notably, a parameter-free BM25 baseline outperformed several off-the-shelf multilingual dense retrieval models on specialist domains.
A research team adapted NVIDIA's Nemotron retrieval stack for Modern Greek, covering corpus mining, synthetic supervision, retrieval training, reranker adaptation, and reader fine-tuning, and introduced a new benchmark called HERA. The work addresses the absence of Modern Greek from Nemotron retrieval models and major multilingual benchmarks, targeting RAG in legal, energy, financial, and medical domains. Notably, a parameter-free BM25 baseline outperformed several off-the-shelf multilingual dense retrieval models on specialist domains.
0 engagement1 sourcearxivWed, Aug 5, 2026, 05:56 PM
Products·launched
Cacophony: compiler for synthetic datasets from YAML schemas
Cacophony is a tool that generates realistic synthetic datasets by treating a dataset as the output of a program defined by a YAML schema. It uses LLMs (built with Claude Code) to interpret the schema and produce coherent records, solving the problem of fake data that lacks relational integrity. It targets developers and data scientists needing realistic test data.
Cacophony is a tool that generates realistic synthetic datasets by treating a dataset as the output of a program defined by a YAML schema. It uses LLMs (built with Claude Code) to interpret the schema and produce coherent records, solving the problem of fake data that lacks relational integrity. It targets developers and data scientists needing realistic test data.
1 engagement1 sourceredditTue, Aug 18, 2026, 09:25 PM
Developments·Paper
Researchers propose two-stage framework for cross-tabular data generation
A new arXiv paper introduces a two-stage framework for cross-tabular data generation (CTDG), addressing the limitation of existing methods that handle only single-input tables. The approach transforms heterogeneous raw tables into standardized statistical tables, enabling generation of synthetic datasets from multiple heterogeneous sources.
A new arXiv paper introduces a two-stage framework for cross-tabular data generation (CTDG), addressing the limitation of existing methods that handle only single-input tables. The approach transforms heterogeneous raw tables into standardized statistical tables, enabling generation of synthetic datasets from multiple heterogeneous sources.
0 engagement1 sourcearxivFri, Aug 14, 2026, 05:10 PM
Releases·Agent Release
CrowdStrike launches SafeMind security AI with Red Tempest and Blue Solano models
CrowdStrike announced SafeMind, an agentic cybersecurity system built with NVIDIA Nemotron and running in the Falcon platform, unveiled by George Kurtz and Jensen Huang at Fal.Con 2026. It pairs an offensive model, Red Tempest, which finds attack paths, with a defensive model, Blue Solano, which closes them. It matters to practitioners because it automates continuous red/blue testing against increasingly automated attacks.
CrowdStrike announced SafeMind, an agentic cybersecurity system built with NVIDIA Nemotron and running in the Falcon platform, unveiled by George Kurtz and Jensen Huang at Fal.Con 2026. It pairs an offensive model, Red Tempest, which finds attack paths, with a defensive model, Blue Solano, which closes them. It matters to practitioners because it automates continuous red/blue testing against increasingly automated attacks.
Updated Sun, Sep 6, 2026, 08:13 AM — WSJ coverage on 2026-09-06 highlighted the partnership's goal of matching AI-driven attack velocity for defenders.
02012 engagement3 sourcesreddit, rssTue, Sep 1, 2026, 07:54 PM