Live · runs on demand

HelaLM speaks Sinhala
like a native.

Most entries adapt older bases like Llama-3.2. We built on Gemma-3-4b, Google's newer-generation open model, and still made it 92.3% better at Sinhala (සිංහල) than itself. Type anything below and watch both models answer.

92.3%win rate vs base
300 prompts, 10 judges
0/300code-switching
(base: 84/300)
+3.6%Belebele
reading comp.
Model· Dataset· Kaggle
Try it, ask in Sinhala

Base gemma-3-4b-it

HelaLM-4B ours

The numbers

Head-to-head win rate Win rate by domain

How it was built, a study not a single run.

We treated this as a controlled experiment: a 4-run ablation isolating the two variables that actually govern low-resource adaptation, before scaling data.

FINDING 01

We picked the harder, better base

The same recipe on the older Llama-3.2-3B produced degenerate Sinhala; on the newer Gemma-3-4b it was 94% preferred (its 262k vocab tokenizes Sinhala far better). Improving a strong modern base is harder than inflating a score off a weak one, and HelaLM still wins 92.3%.

FINDING 02

Targeted data beats more data

A naive larger mixture regressed reading comprehension. A comprehension plus programmatic-math rebalance recovered every benchmark.

METHOD

Blind, 10-judge eval

300 prompts, randomized A/B order, 10 independent judges. Built with Adaptive Data (Adaption Labs); LoRA SFT, r=32, all-linear, 3 epochs.