FPll/Llama-LIMBA-2.0-GGUF

Llama LIMBA 2.0 — GGUF

Built with Llama

Llama LIMBA 2.0 is a fine-tuned Meta Llama 3.1 8B model developed for Italian and Limba Sarda Comuna (LSC).

The project explores the adaptation of a Large Language Model to a minority and low-resource language through parameter-efficient fine-tuning.

Capabilities

Llama LIMBA 2.0 supports:

• Italian → Limba Sarda Comuna translation;
• Limba Sarda Comuna → Italian translation;
• Sardinian text generation;
• conversational interaction;
• educational explanations;
• basic programming assistance;
• experimentation with minority-language NLP.

Model details

Component	Value
Base model	Meta Llama 3.1 8B
Training checkpoint	unsloth/Meta-Llama-3.1-8B-bnb-4bit
Fine-tuning framework	Unsloth
Fine-tuning method	LoRA
Quantized loading	4-bit
Context length	4096 tokens
LoRA rank	32
LoRA alpha	64
Training epochs	3
Export format	GGUF
Quantization	Q4_K_M
Primary language	Limba Sarda Comuna
Secondary language	Italian

Available model file

Meta-Llama-3.1-8B.Q4_K_M.gguf

The GGUF file can be used with compatible runtimes such as llama.cpp and llama-cpp-python.

Interactive demo

LIMBA 2.0 can be tested through its Hugging Face Spaces interface:

https://huggingface.co/spaces/FPll/Limba_2.0

The demo includes a Gradio conversational interface and lightweight retrieval from Sardinian and Italian Wikipedia.

Source code and documentation

The complete project repository includes:

• the public Unsloth and LoRA training notebook;
• the Gradio application;
• local GGUF inference code;
• a curated dataset sample;
• automated dataset validation;
• training methodology;
• qualitative evaluation criteria;
• limitations and responsible-use documentation.

https://github.com/francesco-palladino/LIMBA-Sardinian-Language-Model

Intended use

Llama LIMBA 2.0 is intended for:

• minority-language NLP experimentation;
• Italian–LSC translation assistance;
• educational demonstrations;
• conversational interaction;
• Sardinian text generation;
• research and prototyping;
• testing fine-tuning and deployment workflows.

Limitations

The model may:

• produce grammatically inconsistent or non-standard Sardinian;
• mix LSC, Italian and other Sardinian linguistic forms;
• alter nuances, terminology or tone during translation;
• generate inaccurate, incomplete or unsupported information;
• provide code or educational explanations requiring manual revision;
• perform differently depending on prompt formulation and topic.

The current evaluation process is primarily qualitative and does not yet constitute a formal linguistic benchmark.

Human review is recommended for translations, educational materials, public communications and content requiring linguistic or factual accuracy.

Responsible use

Llama LIMBA 2.0 should not be treated as the sole source for:

• official or certified translations;
• medical, legal or financial decisions;
• emergency information;
• safety-critical applications;
• authoritative linguistic certification.

Generated and retrieval-assisted information should be independently verified.

License and attribution

This model is distributed under the Meta Llama 3.1 Community License.

Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.

Fine-tuning and GGUF conversion were performed using Unsloth.