SGaleshchuk/Llama-2-13b-hf_uk_rank-32_ft

The following training arguments used for Llama-2 finetuning with Ukrainian corpora pf XL-SUM:

• learning-rate=2e-4,
• maximum number of tokens=512,
• 15 epochs. Lora perf arguments:
• rank = 32,
• lora-alpha=16,
• dropout = 0.1.