Skip to main content

Service

LoRA Adaptor Training

Small, swappable adaptors instead of a full copy of the model.

Talk to usSee OpenThai 2.0

OpenThai 2.0 was trained as a rank-64 LoRA on about 143,000 verified rows; its 7 GB adaptor is published.

What we do​

  • Adaptor trainingThe base weights stay frozen; only small low-rank matrices train on your data.
  • One base, many tasksAdaptors load at inference, so one base model serves several tasks.
  • Less GPU memoryOnly the adaptor trains, so training needs less GPU memory than full finetuning.
  • QLoRAQuantised LoRA when GPU memory is the limit.

How we work​

  1. Define the tasks. We set the target tasks and prepare instruction–response pairs.
  2. Train. We choose the rank and target modules for the task, then train the adaptor.
  3. Merge or keep separate. Merged into the base model for one task, or kept apart to swap between tasks.

What you get​

  • The adaptor weights for your base model, merged or separate.

Start​

Tell us the tasks, the data you have and the base model you run, and we reply with a scope and a quote. Email sale@iapp.co.th or call 02-124-4041.