Skip to content

Latest commit

 

History

History
40 lines (33 loc) · 1.34 KB

README.md

File metadata and controls

40 lines (33 loc) · 1.34 KB
Raw

🕴️ finetune LLM 🦙 model on custom DataSet, using unsloth

🕴️ ONESTEP CODE : finetune.py 🦙

  1. Prepare DataSet to match the unsloth format :

process_dataset_to_unsloth.ipynb

Process Dataset -> dataset_to_unsloth.ipynb
  1. Adapt the config arguments : finetune.py in finetune.md

finetune.md

def finetune(
  # -- PARAMETERS CONFIG -- 
  SOURCE_MODEL = "unsloth/Phi-3-mini-4k-instruct",
  DATASET = "0xZee/arxiv-math-Unsloth-tune-50k", 
  #DATASET = "ArtifactAI/arxiv-math-instruct-50k",
  MAX_STEPS = 444,
  FINETUNED_LOCAL_MODEL = "Phi-3-mini_ft_arxiv-math",
  FINETUNED_ONLINE_MODEL = "0xZee/Phi-3-mini_ft_arxiv-math",
  TEST_PROMPT = "Which compound is antiferromagnetic?", # response : common magnetic ordering in various materials.
):
  1. Run the onestep file : finetune.py in finetune.md

finetune.md

python finetune.py

🕴️ finetune llama3.1 🦙 model on custom DataSet

  • 🏬 FineTunning Framework : Unsloth on GPU Tesla T4
  • 🦙 Source Model : models--unsloth--meta-llama-3.1-8b-bnb-4bit Model 🕴️
  • 💾 Training DataSet ; "yahma/alpaca-cleaned" on HuggingFace
  • ⚙️ Fine-Tuned Model : 🕴️ llama3-1_0xZee_model
  • Model saved to : https://huggingface.co/0xZee/llama3-1_0xZee_model