Zolint at WMT 2026: A Multitask LLM Submission for the Low-Resource Ukrainian Track

LoRA adapter for the Ukrainian track of the WMT 2026 Shared Task on Multitask LLMs with Limited Resources (MT, QA, SC, GC, MR).

On the official test set, our unified system obtains the best score on all reported metrics except question answering.

Official test highlights

Training summary

1. Stage 1: LoRA (r=16, α=32) on MT (en→uk) + GC + SC; keep best epoch by combined MT+GC+SC.
2. Stage 2: Continue 1 epoch on QA + MR + MT/GC replay + full Czech→Ukrainian chat data.

Quick start

pip install torch transformers peft accelerate

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen3.5-2B"
lora_id = "MDK-YLC/Qwen3.5-2B-WMT26-Ukrainian-LoRA"

tokenizer = AutoTokenizer.from_pretrained(lora_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, lora_id)
model.eval()

See inference_example.py in this repo for MT / QA / SC-style prompts.

Files

Note: Adapter only — base weights come from Qwen/Qwen3.5-2B.

Citation

Zolint at WMT 2026: A Multitask LLM Submission for the Low-Resource Ukrainian Track
Hugging Face: MDK-YLC/Qwen3.5-2B-WMT26-Ukrainian-LoRA