This will guide you through fine tuning an Open Source LLM using Unsloth

https://docs.unsloth.ai/

https://colab.research.google.com/drive/1oHPzDpApeuGSdOEFribp6Bt85hZphe39#scrollTo=6bZsfBuZDeCL

1. Installation

We will be installing the following packages

!pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install --no-deps "xformers<0.0.27" "trl<0.9.0" peft accelerate bitsandbytes

<aside> ❗ Run pip install -U pyarrow==15.0.2 If pyarrow throws an error for you

</aside>

2. Load the model

from unsloth import FastLanguageModel
import torch

max_seq_length = 2048 # Choose any! We auto support RoPE Scaling internally!
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.

https://huggingface.co/blog/4bit-transformers-bitsandbytes

https://www.youtube.com/watch?v=TPcXVJ1VSRI

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/llama-3-8b-bnb-4bit",
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,
)

We add Lora adapters so we only need to update 1-10% of all parameters

model = FastLanguageModel.get_peft_model(
    model,
    r = 16, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",],
    lora_alpha = 16,
    lora_dropout = 0, # Supports any, but = 0 is optimized
    bias = "none",    # Supports any, but = "none" is optimized
    # [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes!
    use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context
    random_state = 3407,
    use_rslora = False,  # We support rank stabilized LoRA
    loftq_config = None, # And LoftQ
)

3. Data Preparation

Now we will download and prepare our Dataset. We will be using a Medical Q&A dataset

https://www.kaggle.com/datasets/thedevastator/comprehensive-medical-q-a-dataset