Product

News

User Story

Engineering

Culture

Justin Olsson

Waleed Kadous

Llama 2 70b matches legislative interns for summaries, but GPT-4 does better. Studying the GPT-4 summaries helped us improve the prompt. It worked.

Human rated performance of different bill summarization

LLM-based summarization: A case study of human, Llama 2 70b and GPT-4 summarization quality

Kyle Huang

Wendi Ding

Liguang Xie

Avnish Narayan

Ricky Xu

Anyscale is releasing LLMPerf for benchmarking LLMs on current LLM offerings. See benchmarking results for Anyscale Endpoints vs Fireworks.ai.

End-to-end time

Reproducible Performance Metrics for LLM inference

Using Anyscale Endpoints, we compared Llama 2 7b, 13b and 70b vs. OpenAI's GPT-3.5-turbo and GPT-4 for accuracy and cost.

factuality 373 examples

Llama 2 vs. GPT-4: Nearly As Accurate and 30X Cheaper

Llama 2 is about as factually accurate as GPT-4 for summaries and is 30X cheaper

Kourosh Hakhamaneshi

Fine tuning is one approach to domain-specific model refinement (DSMR), but it’s not a silver bullet for improving domain-specific performance.

fine-tune-img-1

Fine Tuning Is For Form, Not Facts | Anyscale

Fine tuning is for form, not facts

Aviary Explorer

Announcing Aviary: Open Source Multi-LLM Serving

For LLM developers, it's useful to know numbers for back-of-the-envelope calculations. Here are the numbers we use, plus how and why to use them.

numbers-cover-image

Numbers Every LLM Developer Should Know | Anyscale

Numbers every LLM Developer should know

In part 3 of our blog series, we’ll show you how to build an LLM question and answering service in 20 minutes using LangChain & Ray.

qna-answering-questions-with

Building a Self-Hosted Question Answering Service in 20 Minutes

Building a Self Hosted Question Answering Service using LangChain + Ray in 20 minutes

In part 1 of a new blog series, we show how to build a search engine in 100 lines using LLM embeddings and a vector database.

Build a document index easily with Ray and Langchain

doc-index-starter-langchain

Building an LLM Open-Source Search Engine in 100 Lines

Building an LLM open source search engine in 100 lines using LangChain and Ray

Jun Gong

Antoni Baum

Richard Liaw

Richard Liaw headshot

In part 4 of our Generative AI series, we share how to build a system for fine-tuning & serving LLMs in 40 minutes or less.

fine-tune-stack

How to Fine-Tune a 6 Billion Parameter LLM for Less Than $7

How to fine tune and serve LLMs simply, quickly and cost effectively using Ray + DeepSpeed + HuggingFace

Eric Liang

Eric Liang headshot

Chen Shen

Clark Zinzow

3rdGenTasks andActors

Data Ingest in a Third-Generation ML Architecture | Anyscale

Deep Dive: Data Ingest in a Third Generation ML Architecture

Uber Leveraging Ray

The 3rd Generation of Production ML Architectures | Anyscale

The Third Generation of Production ML Architectures

WaleedKadous345

Why I’m joining Anyscale

Ray has become the engine for AI-native computing for thousands of AI- and digital-natives. Now Anyscale and Microsoft are partnering to bring the power of Ray to every enterprise.

Announcing first-party service - Anyscale on Azure

Announcing Anyscale Native Offering on Azure: Build, Run and Scale AI-Native Workloads Securely on Azure Infrastructure

Announcing Anyscale First-Party Offering on Azure: Build, Run and Scale AI-Native Workloads Securely on Azure Infrastructure

Ray, the open-source distributed compute framework for AI is joining the PyTorch Foundation, part of The Linux Foundation. 

Ray joins PyTorch Foundation

Ray is Joining The PyTorch Foundation

opensource stack blog thumbnail

Components of an Open Source AI Compute Tech Stack

An Open Source Stack for AI Compute: Kubernetes + Ray + PyTorch + vLLM

Explore a technical comparison of leading Reinforcement Learning (RL) libraries for LLMs from Ray. This guide analyzes frameworks like TRL, Verl, and RAGEN to help developers choose the best tools for RLHF, reasoning, and agentic AI.

The architecture of a Reinforcement Learning (RL) library is split into two primary components: Generation and Training. During the generation phase, an LLM Engine performs multi-turn rollouts within an environment to produce data and reward signals. This output is then fed into the training phase to update the model's parameters. This process forms a feedback loop, where the progressively improved model generates the next iteration of data for continuous refinement.

Open Source RL Libraries for LLMs Graph

Open Source RL Libraries for LLMs

This section is used to order the "Types" and "Tags" that show up for filters on the Blog Index

Types

Products / Libraries

Blog

Join us at Nvidia GTC @ AI Developer Meetup Mar 17.

Powered by Ray, Anyscale empowers AI builders to run and scale all ML and AI workloads on any cloud and on-prem.

Learn why Anyscale is best place to run Ray

Free, self-paced curriculum for practitioners

One-click to launch. Start fast, make it yours

Blogs