Launch an LLM Finetuning Job

❗

Pre-requisites

Before starting, ensure you meet the following requirements:

• MonsterAPI Account: Sign up if you don't have an account.
• API Credits: Minimum of 1,000 credits required. Explore our plans.

Access the Fine-Tuning Portal

Log in to your MonsterAPI account and navigate to the "Fine-Tuning" Portal. Click on "Create New Job."

Step-by-Step Guide

1. Specify Job Name and Select an LLM
   • Enter a unique job name.
   • Choose an LLM model from the drop-down menu based on your needs. Options include Llama 2 7B, CodeLlama, Falcon, GPT-J 6B, and X-Gen.

2. Select a Task and Dataset

• Select a Fine-Tuning Task: Choose from options like "Instruction Fine-Tuning," "Text Classification," or specify a custom task.

• Select a Dataset:
  - Curated Hugging Face Dataset: Choose from our selection. If available, select a subset from the 'Choose a Subset' dropdown.

  

  • Unlisted Hugging Face Dataset: Enter the dataset name for Hugging Face datasets not listed. Configure the prompt in the text section based on your dataset’s columns.

    Example:
    If using the kz919/alpaca dataset:

  ###Instruction:{prompt}
  ###Response:{completion}
  

  • Custom Dataset: Upload and use your dataset. For guidance on uploading, see Custom Datasets. Choose your dataset from "My Datasets" if already uploaded.

  

3. Specify Hyperparameter Configuration

• Set hyperparameters like epochs, learning rate, cutoff length, and warmup steps. Default values are provided based on the selected model but can be adjusted as needed.

• Note: Cutoff length affects batch size and speed but does not impact the model’s context length, which is determined by the base model.

  
  

4. Review and Submit Job

• Click "Next" to proceed to the summary page.

• Review all settings and submit your job. The finetuning process will start shortly, and you can monitor job logs and metrics once it is in progress.

⚙️ Optional Settings

1. Track with WandB

   • Add your WandB credentials:
     • Username
     • API Key (Get it here)
     • Project Name
     • Run Name

   Metrics will be sent to your WandB project for tracking.

2. Upload to Hugging Face

   • Add Hugging Face credentials:
     • API Key (Must have write access)
     • Repo Path

The finetuned model will be published to your Hugging Face repository upon completion.