choosing an LLM for a translation project can be confusing.
T5, Flan-T5, Qwen... there are so many open-source options. But how do you know which one is right for your task?
I had a Kinyarwanda–English translation project that wasn't working. I trained a tiny GPT-style model from scratch, and it produced broken sentences like "this is the was to see this" instead of "this door is black."
I needed a real LLM. I started by comparing a few suitable models and looking at four things:
Architecture: Does the model fit my type of task?
Translation needs an encoder-decoder. T5 and Flan-T5 are built for this. Qwen is decoder-only, built for chat and generation.
Hardware: Does it run on a free Google Colab T4 (16 GB)?
Both T5 and Flan-T5 need less than 1 GB. Qwen needs about 9–12 GB. All three fit, but the smaller models train much faster.
Access: Is it free without an API key?
T5, Flan-T5, and Qwen are all Apache 2.0. No account creation, no hidden fees.
Language Support: Does it know my language?
Here is the hard truth: none of them were pretrained on Kinyarwanda. I tested Flan-T5 before fine-tuning, and it gave me "a syringe" for "this door is black."
I wouldn't choose a model just because it's popular or has billions of parameters. I picked Flan-T5. It's an encoder-decoder designed for translation, it’s small enough to run fast on a free T4, and its instruction tuning means it already understands the "translate this" prompt format. I just need to fine-tune it on my 3,000 sentence pairs.
💡 My takeaway: For low-resource tasks, the best model isn't the biggest chat model. It's the one with the right architecture that fits your hardware.
A question for experienced ML engineers: When you have a small dataset, do you prefer fine-tuning a smaller, task-specific model or trying to prompt a larger one?
Open to learning from others in the community!
Top comments (0)