# LLaMA 3.2 3B — Python Code Generation (QLoRA Fine-tune) ## Model Description Fine-tuned LLaMA 3.2 3B on Python code generation using QLoRA (4-bit NF4 quantization). Trained on `iamtarun/python_code_instructions_18k_alpaca` — 18k Python instruction-output pairs. ## Training Details - **Base model:** meta-llama/Llama-3.2-3B - **Method:** QLoRA (LoRA rank=16, alpha=32, 4-bit NF4 quantization) - **Target modules:** q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj - **Dataset:** 16,738 training samples, 1,862 test samples - **Epochs:** 3 - **Batch size:** 2 × 8 gradient accumulation = 16 effective - **Learning rate:** 2e-4 with linear scheduler - **Optimizer:** AdamW 8-bit - **Hardware:** Single NVIDIA T4 (16GB) - **Training time:** ~3.6 hours - **Final training loss:** 0.573 ## Loss Curve | Epoch | Loss | |---|---| | 1 | ~0.65 | | 2 | ~0.47 | | 3 | 0.57 | ## Usage ```python from unsloth import FastLanguageModel model, tokenizer = FastLanguageModel.from_pretrained( "reddyrohith49471/llama3.2-3b-python-qlora", max_seq_length=2048, load_in_4bit=True, ) FastLanguageModel.for_inference(model) inputs = tokenizer( "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n" "Write a Python function to check if a number is prime." "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", return_tensors="pt" ).to("cuda") outputs = model.generate(**inputs, max_new_tokens=256) print(tokenizer.decode(outputs[0], skip_special_tokens=True)) ``` ## Framework - Unsloth 2x faster fine-tuning - HuggingFace PEFT + TRL - PyTorch + bitsandbytes