Add model card with exact and within-1 confusion matrices and per-class metrics

Upload fine-tuned LLaMA-3-8B CEFR model with exact and within-1 confusion matrices

initial commit