llama31-8b-finetune-multit

Llama 3.1 8B Ita — Italian Cultural Alignment [V1]

Llama 3.1 8B Ita [V1] is a LoRA adapter fine-tuned on top of DeepMount00/Llama-3.1-8b-ITA to improve Italian cultural alignment. It was trained on the Mult-IT dataset and evaluated on the ITALIC benchmark. Unlike Qwen3, Llama 3.1 is a standard causal language model without a hybrid reasoning architecture, so no thinking-mode considerations apply.

Author: Maruf Bepary, King's College London
Research report: Alignment in Large Language Models

---

Model Summary

---

Intended Use

This model is intended for:

• Italian language understanding — multiple-choice Q&A, cultural knowledge, and general instruction following in Italian.
• Research — comparing the effect of SFT on Italian cultural alignment across model families.
• Benchmarking — comparing Italian-specific models against multilingual and fine-tuned baselines.

Not recommended for:

• High-stakes or safety-critical applications.
• Languages other than Italian.

---

Key Finding — Cultural Alignment

Training on the Italian cultural Q&A dataset (Mult-IT) improves performance across almost all ITALIC categories:

Language competence improved more than culture knowledge. The largest gains were in Synonyms (+8.76 pp), Morphology (+8.29 pp), Orthography (+7.03 pp), and Civic (+6.07 pp). Events remained flat (0.00 pp change). As Llama 3.1 does not have a hybrid reasoning architecture, fine-tuning carries no risk of reasoning-mode degradation.

---

Training Details

LoRA Configuration

Training Hyperparameters

Note: full training hyperparameters are detailed in the research report.

Framework & Hardware

Training Dataset — Mult-IT

• Dataset: Mult-IT — Multiple Choice Questions on Multiple Topics in Italian
• Source: CALAMITA Shared Task @ CLiC-it 2024
• Language: Italian
• Size: ~86,929 training samples
• Format: JSONL, multiple-choice Q&A
• Reference: Mult-IT: Multiple Choice Questions on Multiple Topics in Italian (2024)

---

ITALIC Benchmark Results

Benchmark: ITALIC (NAACL 2025) — Italian Culture-Aware Natural Language Benchmark
Format: Zero-shot, multiple-choice (12 categories, 10,000 questions)
System prompt: "Sei un assistente utile."

V1 vs Baseline

Comparison with Other Models (ITALIC Total)

All scores evaluated under identical zero-shot conditions on the ITALIC benchmark.

---

Usage

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

base_model_id = "DeepMount00/Llama-3.1-8b-ITA"
adapter_id = "maruf-bepary/llama-3.1-8b-ita-italian-v1"

# Load tokeniser and base model
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

# Load LoRA adapter
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

# Example: Italian multiple-choice question
messages = [
    {"role": "system", "content": "Sei un assistente utile."},
    {
        "role": "user",
        "content": (
            "Qual è la capitale d'Italia?\n"
            "A) Milano\nB) Roma\nC) Napoli\nD) Torino\n\n"
            "Rispondi con la lettera della risposta corretta."
        ),
    },
]

# Apply LLaMA-3 chat template
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=64,
        do_sample=False,
        temperature=None,
        top_p=None,
    )

response = tokenizer.decode(
    outputs[0][inputs["input_ids"].shape[-1]:],
    skip_special_tokens=True,
)
print(response)
# Expected output: "B"

---

Limitations

• Morphology (60.43%) and Syntax (54.31%) remain the weakest categories despite improvement.
• Benchmark scope — evaluation was conducted solely on ITALIC; performance on other Italian benchmarks is unverified.
• Single-GPU training — training used one RTX 3090; multi-GPU configurations may yield different results.
• Dataset bias — Mult-IT is a multiple-choice dataset; generalisation to open-ended Italian generation tasks is unverified.
• Events category showed no improvement (0.00 pp), suggesting the training data may lack current-events coverage.

---

References

Related resources:

• Research report: Alignment in Large Language Models
• Base model: DeepMount00/Llama-3.1-8b-ITA
• ITALIC benchmark: RiTA-nlp/ITALIC
• Mult-IT dataset: sapienzanlp/Mult-IT
• PEFT documentation: huggingface.co/docs/peft