21 articles
Hugging Face deploys its text-generation pipeline on Intel Gaudi 2 AI accelerator. The solution optimizes language model inference on Intel hardware, providing an alternative to traditional GPUs for production deployments.
Hugging Face releases StarCoder2, open-source code model trained on The Stack v2 (17B code tokens dataset). StarCoder2 outperforms CodeLlama on HumanEval and MBPP benchmarks. Available in 3B, 7B, and 15B variants with open weights and source code.
Hugging Face launches TTS Arena, a benchmarking platform to evaluate text-to-speech models in real-world conditions. Users compare audio outputs from different TTS models through voting, creating a dynamic ranking based on human preferences.
Hugging Face publishes a guide on AI watermarking techniques, covering tools and methods for marking AI-generated content. The guide explores existing approaches and use cases for tracing synthetic data origins.
Hugging Face releases comprehensive guide for fine-tuning Gemma models. Documentation covers memory optimization techniques, LoRA usage, and recommended configurations for adapting Gemma to specific tasks.
Hugging Face introduces Matryoshka embedding models, enabling dynamic dimensionality reduction of vectors without retraining. This approach provides flexibility and efficiency for RAG and vector search applications.
Hugging Face introduces a red-teaming resistance leaderboard to evaluate AI models' robustness against adversarial attacks. The initiative measures systems' ability to withstand attempts to bypass safety guardrails.
Google releases Gemma, a new family of open-source language models based on Gemini architecture. Available in 2B and 7B versions, Gemma is optimized for efficiency and runs on consumer hardware.
Hugging Face introduces the Open Ko-LLM Leaderboard, a ranking system dedicated to evaluating Korean language models. This initiative establishes a standardized evaluation ecosystem for Korean LLMs, enabling performance comparison across benchmarks tailored to Korean linguistic context.
PEFT adds new model merging methods. These techniques enable efficient combination of multiple fine-tuned adapters without significantly increasing model size.
Hugging Face explores synthetic data to reduce costs, time, and carbon footprint in AI training. The open source approach enables generating training data without large-scale real data collection, optimizing resources and sustainability.
OpenAI introduces Sora, a text-conditional diffusion model trained jointly on videos and images of variable durations, resolutions and aspect ratios. Built on a transformer architecture operating on spacetime patches, Sora generates up to one minute of high-fidelity video. OpenAI suggests that scaling video generation models is a promising path toward general-purpose physical world simulators.
OpenAI identified and disabled accounts used by state-affiliated threat actors for disinformation, espionage, and malicious capability development operations. The company publishes detailed reports on these campaigns and strengthens detection measures.
AMD launches a developer contest for pervasive AI applications. The competition rewards solutions leveraging AMD processors for model inference and training.
OpenAI is testing memory functionality for ChatGPT to retain information from previous conversations. Users have full control over memory activation and management.
Hugging Face releases a Messages API compatible with OpenAI for open-source models. The interface unifies access to Claude, Llama, Mistral and other LLMs through a standardized endpoint, reducing friction for migration from OpenAI.
Segmind introduces SegMoE, a diffusion model using a Mixture of Experts architecture for image generation. The system combines multiple specialized experts to improve quality and diversity of generated outputs.
Hugging Face releases NPHardEval Leaderboard, a benchmark assessing LLM reasoning abilities through NP-hard problems and dynamic updates. The leaderboard ranks models by performance on tasks of increasing complexity.
Hugging Face deploys Text Generation Inference (TGI) on AWS Inferentia2. This integration optimizes language model inference on Amazon's specialized hardware, reducing latency and costs for production deployments.
Hugging Face integrates Patch Time Series Transformer, a model specialized in time series forecasting. The implementation provides an alternative to traditional approaches with a transformer architecture adapted for temporal data.
Hugging Face introduces Constitutional AI approach for open-source LLMs. The method applies defined ethical principles to guide model training and evaluation, independent of proprietary APIs.