GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
MDK-YLC
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
MDK-YLC
Model Tree
Base
Qwen/Qwen3.5-2B
Adapter
this model
Input Modalities
Text
Image
Video
Output Modalities
Text
Supported Functionality
Dedicated Endpoints
| Item | Value |
|---|---|
| Team | Zolint |
| Institution | UTokyo & TUM |
| Base model | Qwen/Qwen3.5-2B |
| Adapter type | PEFT LoRA |
LoRA rank (r) | 16 |
| LoRA alpha | 32 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Training | Two-stage SFT (Stage1: MT+GC+SC → Stage2: QA+MR + MT/GC/MT_CS replay) |
| Task | Metric | Score |
|---|---|---|
| MT | BLEU | 18.72 |
| MT_CS | BLEU | 21.45 |
| QA | accuracy | 38.2% |
| SC | detection / correction | 76.1% / 61.4% |
| GC | detection / correction | 38.2% / 37.8% |
| MR | accuracy | 25.0% |
bash
pip install torch transformers peft accelerate
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModelbase_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()messages = [{"role": "system", "content": "You are a professional English-to-Ukrainian translator."},{"role": "user", "content": "Hello, how are you?\n\nUkrainian:"},]prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(prompt, return_tensors="pt").to(model.device)with torch.no_grad():out = model.generate(**inputs, max_new_tokens=128, do_sample=False)print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
See inference_example.py in this repo for a slightly fuller example covering MT / QA / SC-style prompts.
| File | Description |
|---|---|
adapter_config.json | LoRA / PEFT config |
adapter_model.safetensors | LoRA weights |
tokenizer.json / tokenizer_config.json | Tokenizer used at train / test time |
chat_template.jinja | Chat template |
inference_example.py | Minimal runnable example |
Note: This repo contains the LoRA adapter only. You still need the base model
Qwen/Qwen3.5-2B (downloaded automatically by from_pretrained above).
If you use this adapter, please cite the WMT26 Limited-Resources shared task and the Qwen3.5 model card.