BECUAI

Product

[RDPLINE Use Case] Building a Domain-Specific Korean Corpus for a Foundation Model

2026.11.13


domain-specific-training-corpus.jpg
What separates a general model from one that is strong in expert fields comes down to what it was trained on. To handle medical or financial knowledge well in Korean, a model needs Korean text from those fields at real density.

A Korean enterprise AI lab building its own foundation model came to BECUAI for exactly that: a training corpus focused on expert domains.

The challenge: data that fits both the rights and the domain

First, news scraped from the web isn't rights-cleared, so it can't be used for commercial training. Second, even licensed news won't move domain performance much if it's mostly general news. And signing specialist outlets one by one means different terms and formats every time, which drives up coordination costs.

AI training adds its own questions on top:

  • Can the trained model quote parts of the original text?
  • Who owns the model and its outputs?
  • What happens to an already-trained model when the data contract ends?

Solution 1: grouped by domain, not by outlet

  • Domain-specific outlet mix: trade magazines and specialist online news from fields such as healthcare, finance, and IT were combined into one corpus.
  • Category-level filtering: for outlets with broader coverage, politics, society, and culture articles were removed, leaving only industry and economy coverage. Domain fit was managed at the category level, not the outlet level.
  • Adjusted before delivery: as the lab's training plans firmed up after signing, the outlet mix and volume were revised, and the full set was delivered as revised.

That last point matters more than it seems. Training plans usually take shape after the contract is signed. Locking the data mix at signing makes it hard to reflect what the model actually turns out to need.

Solution 2: rights designed to outlast the model

  • Verified rights: only data lawfully obtained from rights holders was included, with no scraped data. Proof of license was submitted at delivery.
  • Scope of use built for training: model training, partial quotation of the original, and publishing transformed outputs are allowed. Only external disclosure and redistribution of the data itself are prohibited.
  • Model rights kept separate: when the data term ends, the client keeps its rights to the models trained on this data up to that point, and to their outputs.

rdpline-usecase-domain-corpus-structure-en.jpg

The outcome

The corpus was used to train the client's foundation model, which has since been improved over several generations. Model performance reflects many datasets and the lab's own research. The BECUAI corpus contributed the Korean text for expert domains. As the model moved from one generation to the next, there was no need to renegotiate data rights.

The takeaway

If you're building a model for expert domains, design the training data around those domains. And look beyond volume and domain fit: check what happens to your model after the data contract ends.

Building a Korean model for a specific field? Let's start with the right mix of outlets for your domain.

Try the RDPLINE API