SemanticAlignment/Llama-3.1-8B-Italian-LAPT-instruct
TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kPublished:Feb 9, 2026Architecture:Transformer Cold
Llama-3.1-8B-Italian-LAPT-instruct is an 8 billion parameter instruction-tuned Llama 3.1 model developed by SapienzaNLP, ISTI-CNR, and ILC-CNR. This model is continually trained and instruction-tuned, specifically adapted for the Italian language. It demonstrates enhanced performance on Italian benchmarks like ITA-Bench, making it suitable for Italian-centric natural language processing tasks.
Loading preview...
Overview
Llama-3.1-8B-Italian-LAPT-instruct is an 8 billion parameter large language model, part of the Llama-3.1-8B-Adapted collection. Developed by SapienzaNLP, ISTI-CNR, and ILC-CNR, this model is a continually trained and instruction-tuned variant of the Llama 3.1 architecture, specifically optimized for the Italian language.
Key Capabilities & Training
- Italian Language Adaptation: The model was adapted using a custom dataset skewed towards Italian, combining 9 billion tokens from the Italian part of CulturaX with 3 billion English tokens.
- Instruction Tuning: Further fine-tuned on a diverse set of instruction-following datasets, including Italian and multilingual resources like TÜLU-v3, LIMA, WildChat-IT, TowerBlocks-v0.2, GPT-4o-ITA-Instruct, and Aya.
- Performance: Achieves competitive results on the ITA-Bench evaluation suite, scoring 58.5 on MMLU (5-shots), 47.9 on ARC-C (5-shots), 62.4 on Hellaswag (0-shots), and 67.3 on IFEval (inst_level), outperforming the original Llama-3.1-Original and Mistral-0.1 models on these Italian-focused metrics.
Good For
- Applications requiring strong performance in Italian natural language understanding and generation.
- Instruction-following tasks in Italian.
- Research and development focusing on multilingual LLMs with a specific emphasis on Italian.