Fine-tuning on Lamini

#2
by zaaath - opened

Hi @sudocoder and the Lamini team,

I'm trying to replicate fine-tuning of EleutherAI/pythia-70m to get something like this model lamini/lamini_docs_finetuned. I'm using the dataset from Sharon's DeepLearning.AI course (see lamini_docs.jsonl in the Jupyter notebook for https://learn.deeplearning.ai/finetuning-large-language-models/lesson/5/data-preparation) and the Lamini platform for this. However, I'm getting poor results, fine-tuning EleutherAI/pythia-410m doesn't seem to converge, while meta-llama/Llama-2-7b-hf is converging. Not sure if I'm missing something here.

I'd really appreciate it if you guys could help me understand the following:

  1. Was the same training dataset used here as in the DeepLearning.ai course?
  2. Is EleutherAI/pythia-70m the base model?
  3. How many epochs did it take to get the fine-tuned model?

Thank you!

P.S. Manual evaluation of the fine-tuned models using "What is Lamini value prop?":

  1. lamini/lamini_docs_finetuned:
    "Lamini is a Python package that provides a base model for generating text in a specific format. It provides a method for generating text in a specific format, such as a word or a word of a specific kind, and can be used to generate text that is representative of the specific kind. Additionally, Lamini provides a method for generating text in a specific format, such as a word or a specific kind, and can be used to generate text that"
  2. Fine-tuned EleutherAI/pythia-410m:
    "Is it a library or a tool? Is it a tool for developers to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use? Is it a library for users to use"
  3. Fine-tuned meta-llama/Llama-2-7b-hf:
    "Lamini is a language model that can generate text based on a given prompt. Can you explain what the code does and how it works?Lamini is a language model that can generate text based on a given prompt. The code explains the setup of the program and how to use it. It includes importing necessary modules, defining variables, and creating a function that takes in a prompt and returns a generated story. The code also includes examples of prompts and generated stories to demonstrate the program's capabilities. Overall, the code provides a framework for generating stories using Lamini."
  4. Fine-tuned gpt-3.5-turbo-1106:
    "Lamini’s value proposition lies in its provision of a distinct engine that is optimized to process language-based algorithms. Efficiencies gained through this engine result in more time-effective development and deployment of language models."

Sign up or log in to comment