llama3.2-finetuned-unsloth

🦙 Fine-tuned LLaMA 3.2 3B (LoRA Adapter via Unsloth)

This repository contains LoRA adapter weights for the unsloth/Llama-3.2-3B-Instruct model, fine-tuned using the Unsloth library in Colab.

📊 Fine-Tuning Details

• Base Model: unsloth/Llama-3.2-3B-Instruct
• Dataset: mlabonne/FineTome-100k
• Fine-Tuning Method: LoRA / PEFT
• LoRA Rank (r): 16
• Target Modules: q_proj, v_proj, o_proj, gate_proj, down_proj
• Sequence Length: 2048 tokens
• Training Steps: 60
• Batch Size: 2 (with gradient_accumulation_steps=4)
• Precision: Mixed (fp16 or bf16, depending on GPU)

🛠 How to Use

from unsloth import FastLanguageModel

# Load base model
base_model, tokenizer = FastLanguageModel.from_pretrained("unsloth/Llama-3.2-3B-Instruct")

# Load LoRA adapter
model = FastLanguageModel.load_peft_weights(base_model, "your-username/llama3.2-finetuned-unsloth")

This model uses the LLaMA 3.1 chat template applied via:

tokenizer = get_chat_template(tokenizer, chat_template="llama-3.1")