DanieleMarcoaldi/EuroLLM-22B-Instruct-LoRA50

EuroLLM-22B Legislative Text — LoRA 50% Merged (AI4TRAD)

Merged version of a LoRA fine-tuned EuroLLM-22B-Instruct-2512, with LoRA weights scaled at 50% before merging. This scaling balances the fine-tuned specialization with the base model's instruction-following capabilities.

Model Details

Field	Value
Base model	utter-project/EuroLLM-22B-Instruct-2512
Fine-tuning method	LoRA (Low-Rank Adaptation), merged at 50% scaling
LoRA rank	32
LoRA alpha	64
LoRA target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Task	EN → 23 EU languages translation (legislative domain)
Precision	bf16
Developed by	AI4TRAD
Organization	European Parliament
Compute	EuroHPC (Discoverer supercomputer, Sofia, Bulgaria)

Why LoRA at 50%?

Applying LoRA weights at 50% (instead of 100%) produces better results in our evaluations: the model retains the base model's general instruction-following and fluency while incorporating the domain specialization from fine-tuning. This is equivalent to setting scaling = 0.5 * (alpha / rank) during inference.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_path = "DanieleMarcoaldi/EuroLLM-22B-Instruct-LoRA50"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(
    model_path, torch_dtype=torch.bfloat16, device_map="auto"
)

messages = [
    {"role": "user", "content": "Translate the following English text to French:\n\nThe regulation applies to all member states."}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Authors

Daniele Marcoaldi

Acknowledgements

This work was supported by EuroHPC resources on the Discoverer supercomputer (Sofia, Bulgaria).