Skip to content

ALMA-13B-R

A translation model fine-tuned with CPO and LoRA

Visit official site
Category AI models

What ALMA-13B-R is

The model is described as a machine translation model that builds on ALMA and applies LoRA fine-tuning with Contrastive Preference Optimization, which the material says can match or surpass GPT-4 or WMT winners.

What you can do with it

  • Download the ALMA and ALMA-R models
  • Download the matching datasets
  • Fine-tune and deploy the model

Model notes

  • CPO takes the place of supervised fine-tuning used in ALMA
  • Triplet preference data is needed for preference learning
  • The model is hosted on a Hugging Face page

Pros & cons

✓ What we like

  • Model files and datasets are available to download
  • Fine-tuning and deployment are named as supported uses

! What to watch out for

  • CPO fine-tuning requires triplet preference data
  • The comparison with GPT-4 and WMT winners is stated in the source material

FAQ

How is it tuned?

The material describes LoRA fine-tuning with Contrastive Preference Optimization instead of supervised fine-tuning.

What can I download?

The ALMA and ALMA-R models plus their datasets from a GitHub repository.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review