Overview

Radu1999/Mistral-Instruct-Ukrainian-SFT is an instruction-tuned language model built upon the Mistral-7B-v0.2 architecture. Developed by Radu Chivereanu, this model focuses on providing strong performance for tasks in the Ukrainian language through supervised fine-tuning.

Key Capabilities

• Ukrainian Language Proficiency: Specifically fine-tuned on a diverse set of Ukrainian datasets, including UA-SQUAD, Ukrainian StackExchange, and a Ukrainian subset of the Belebele Dataset.
• Instruction Following: Designed to respond effectively to instructions, leveraging the [INST] and [/INST] token format for prompts.
• Mistral Architecture Benefits: Inherits architectural features from Mistral-7B-v0.2, such as Grouped-Query Attention, Sliding-Window Attention, and a Byte-fallback BPE tokenizer, contributing to efficient processing.

Performance

Evaluations on the Open LLM Leaderboard show competitive performance for a specialized model:

• Average Score: 62.17
• HellaSwag (10-Shot): 83.12
• MMLU (5-Shot): 60.95
• Winogrande (5-shot): 77.51

Good For

• Applications requiring instruction-based text generation and understanding in Ukrainian.
• Researchers and developers working on Ukrainian NLP tasks who need a specialized model.