From the source

Putting RL back in RLHF

Developer tool

Image published with the post by Hugging Face — Image from huggingface.co

Hugging Face introduces the RLOO (REINFORCE Leave One-Out) Trainer in TRL, a new online RLHF training algorithm that is more memory-efficient and faster than PPO.

From the source
We are excited to introduce the RLOO (REINFORCE Leave One-Out) Trainer in TRL.
huggingface.co

Other stories

Hugging Face· 16 OctNew model

Welcome Beam: a 501B open model you can own, customize, and scale

Reflection's 501B MoE model is open under Apache 2.0

Hugging Face· 15 OctModel update

Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint

IBM's Granite Switch 4.1 models combine chat, decisions, and speech in one endpoint

Announced: 12 Jun 2024
Detected: 03 Aug 2026 21:12 UTC (+781d after)

Visit source ↗huggingface.co

Hugging Face· 15 OctModel update

Apertus 1.5 Release Notes

Open-weight models gain vision, audio, and thinking mode

Hugging Face· 15 OctOpen-weight release

Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta

Meta releases open-weight multimodal MoE model with 816B parameters

Hugging Face· 15 OctOpen-weight release

Welcome FLUX.3 Image Klein in Diffusers

Open image model adds layout control via bounding boxes

Hugging Face· 8 Oct

The model that didn't exist, so you made it yourself

A user creates six custom models via ML Intern in days