You have three LoRA adapters for three tasks on the same base model. Would you merge them, serve them separately, or train one multi-task model?
Prompting, RAG or fine-tuning · Pro 2
Short answer
It depends on deployment and precision. If adapters can share one base cheaply, separate adapters keep each task at its best. If I need one model on limited hardware, I would try merging, first a weighted average, then TIES or DARE, testing each task. With enough data, one multi-task fine-tune is also an option.
Deeper answer
Merging avoids the forgetting of sequential training and needs no GPUs, but precise tasks such as extraction often suffer most. I would compare every option on each task's own test set, plus general and safety checks.
A mixed plan is common: merge the tasks that tolerate it and keep the most precise one as a separate adapter. Whatever I choose, I would version the inputs and settings so the merge can be rebuilt when an adapter changes.
What a strong answer covers
- Considers deployment constraints
- Separate adapters keep each task at its best
- Tries merging methods in order
- Tests per task and considers a mixed plan
Answer it out loud in the mock interview about 1.5 min
Follow-ups to expect
- Which task would you expect to suffer most from merging?
- How would you serve many adapters efficiently?
Common mistakes
- Merging without per-task tests
- Assuming one approach fits all tasks
Learn it
Model merging: combining fine-tuned models into one
- How would you build a dataset to fine-tune an LLM for a business task?
- Your fine-tuned model is no better than the base model. How do you debug it?
- Models now have very large context windows. Why not put all the documents in the prompt instead of using RAG?
- How do you decide which model size to use for a feature?
- How would you choose a base model to continue training or fine-tune for a specialised domain or language?
- What is prompt caching, and when does it help?
- When would you use prompting, RAG or fine-tuning?
- How would you build an AI assistant that works well in several Indian languages?
- What is model distillation, and when would you use it?
- When is fine-tuning worth the effort?
- A reasoning model is more accurate on your task but five times the cost and much slower. How do you decide whether to use it?
- When would you use synthetic data to train or test a model, and what are the risks?