Fine-Tuned LLaMA-3-8B CEFR Model

• Base Model: unsloth/llama-3-8b-instruct-bnb-4bit
• Fine-Tuning: LoRA with SMOTE-balanced dataset
• Training Details:
  • Dataset: CEFR-level sentences with SMOTE and undersampling for balance
  • LoRA Parameters: r=32, lora_alpha=32, lora_dropout=0.5
  • Training Args: learning_rate=2e-5, batch_size=8, epochs=0.1, cosine scheduler
  • Optimizer: adamw_8bit
  • Early Stopping: Patience=3, threshold=0.01
• Evaluation Metrics:
  • CEFR Classifier Accuracy: 0.167
  • Precision (Macro): 0.042
  • Recall (Macro): 0.167
  • F1-Score (Macro): 0.067
  • Perplexity: 14.218
  • Diversity (Unique Sentences): 1.000
  • Inference Time (ms): 2216.789
  • Model Size (GB): 4.8
  • Robustness (F1): 0.063
• Usage:
  python

  1from transformers import AutoModelForCausalLM, AutoTokenizer 2 3model = AutoModelForCausalLM.from_pretrained("Mr-FineTuner/Test___01") 4tokenizer = AutoTokenizer.from_pretrained("Mr-FineTuner/Test___01") 5 6# Example inference 7prompt = "<|user|>Generate a CEFR B1 level sentence.<|end|>" 8inputs = tokenizer(prompt, return_tensors="pt") 9outputs = model.generate(**inputs, max_length=50) 10print(tokenizer.decode(outputs[0], skip_special_tokens=True))