HiTZ/Latxa-Llama-3.1-70B-Instruct-v2
HiTZ/Latxa-Llama-3.1-70B-Instruct-v2 is a 70 billion parameter instruction-tuned language model developed by HiTZ Research Center & IXA Research group (University of the Basque Country UPV/EHU). Based on Meta's Llama-3.1-Instruct, it has been further trained on an extended Basque corpus (Latxa Corpus v1.1) to significantly enhance performance in Basque language tasks and chat conversations. This model excels in low-resource language scenarios, outperforming its parent model on Basque benchmarks and ranking competitively against proprietary models like GPT-4o and Claude Sonnet in arena-based evaluations for Basque.
Loading preview...
Model Overview
HiTZ/Latxa-Llama-3.1-70B-Instruct-v2 is a 70 billion parameter instruction-tuned Large Language Model (LLM) developed by the HiTZ Research Center & IXA Research group. It is built upon Meta's Llama-3.1-Instruct architecture and has been specifically fine-tuned on an extensive Basque corpus (Latxa Corpus v1.1).
Key Capabilities
- Enhanced Basque Language Performance: Significantly outperforms the base Llama-3.1-Instruct model on standard Basque benchmarks and in Basque chat conversations.
- Low-Resource Language Specialization: Addresses the performance gap for low-resource languages like Basque, promoting LLM-based technology development in these contexts.
- Instruction Following: Trained to effectively follow instructions and function as a chat assistant in Basque.
- Competitive Performance: Achieved 3rd place in a public arena-based evaluation against models like GPT-4o and Claude Sonnet, surpassing other same-size competitors.
Evaluation Highlights
Latxa 3.1 70B Instruct V2 demonstrates strong results across various Basque-specific datasets, including Belebele, X-StoryCloze, EusProficiency, EusTrivia, and EusExams. For instance, it scored 73.30% on EusProficiency and 75.10% on EusTrivia, showing substantial improvements over the original Llama-3.1 70B Instruct.
Intended Use
This model is primarily intended for use with Basque language data for instruction following and chat applications. It inherits the Llama-3.1 license, permitting both commercial and research use.