title: Choctaw Nation and Oracle
aliases: [Choctaw, Choctaw Nation LLM]
tags: [example, case]
summary: A Nation-led collaboration with a commercial vendor to build a Choctaw translation model, with the language data kept under Choctaw control.
genre: Case
themes: [Data sovereignty, Digitization and legacy materials, Capacity-building]
technology: [LLM, MT]
stance: Promising

Choctaw Nation and Oracle

The Choctaw Nation of Oklahoma worked with Oracle to build a custom system for translating between English and Choctaw, building on a long-standing relationship between the Nation and the vendor. With only a few hundred first-language speakers remaining, the Nation wanted to move beyond a fully manual translation process, and it approached the work as its own project with a vendor engaged to help, rather than as a vendor's project done to a community.

The first thing that makes this interesting is the data sovereignty arrangement: sensitive language materials stay inside Choctaw-controlled infrastructure rather than being sent off to a third-party service, which addresses the concern that keeps many communities away from commercial AI. The second is the technical approach. The system fine-tunes an open-weight model (Llama) rather than calling out to a closed-weight commercial one, and pairs it with retrieval so that translations are grounded in the Nation's own vetted material. It also does real work on legacy materials, extracting and validating translation pairs from difficult sources like PDFs of 19th-century newspapers and interview transcripts carrying unique diacritics, a job that would be impractical to do by hand at scale.

Upsides

• The community set the terms. The Nation drove the project and the vendor built to its requirements, on top of an established relationship rather than a cold approach. This is close to what workshop participants described as the collaboration they actually want from industry: a finished, working product built with language-specific development, not a proof of concept.
• Data stays under community control. Language material is not handed to an outside service to be absorbed into someone else's training data.
• Open-weight foundation. Fine-tuning an open-weight model means the resulting weights are something the Nation can hold, inspect, and keep, which is a very different position from renting access to a closed model.
• Legacy material put to use. The pipeline turns hard-to-read historical documents into usable, validated translation pairs.

Open questions

• Sustainability and lock-in. The system is built on a commercial cloud stack, which carries ongoing cost and vendor dependence. What happens to the system if the relationship or the pricing changes is a fair question for any community considering this pattern.
• Sourcing. Most public detail comes from the vendor's own case study, which is marketing material. Independent description of how well the translation actually works, and how output is reviewed by speakers, would strengthen it.