Instruction Finetuning: Unlock the Power of Large Language Models

Abhijeet Kumar (~abhijeet3922)




Large Language models (LLMs) are powerful tools that can be used for variety of tasks such as question answering, translation, content generation and other natural language understanding tasks. However, there are few challenges currently with open source LLMs trained on massive internet data. Two such reasons are:

  • LLMs may not be accurate for complex tasks.
  • LLMs may not have learnt nuances of specific domain well hence does not hold good for such in-domain tasks.

These issues can be resolved by better aligning LLMs by Instruction Finetuning techniques.

Why does it matter ? Many businesses or captive companies intends to owning their models, and create superior quality models for their domain specific applications without handing their sensitive data over to third parties.


The workshop will cover the mainly the following sections.

1. Prompting LLMs (zero/few shot) 
2. Few-shot Finetuning LLMs
3. Instruction Finetuning of LLMs

The workshop may be designed that includes both a tutorial on implementing Instruction Finetuning techniques and hands-on practice session where participants can apply the same on other tasks.

We will cover these section using HuggingFace and other useful libraries for the above agenda. Participants can use Google Collab to perform all the above tasks. For audience aware of LLMs, the workshop intends to demonstrate "How to train your own Alpaca or Dolly models using Llama-2 models?". I will begin by covering the basics of working with LLMs and some of few-shot finetuning techniques. I will then build on this foundation to demonstrate how to train replicable models like Alpaca, Dolly etc. with prompt-response pairs.

Audience This workshop is intended for any developers who are interested in finetuning LLMs for improved accuracy. No prior experience of Finetuning LLMs is required. Basic to Intermediate skillset should be fine.

Outcomes By the end of the workshop, participants will be able to:

  • Working with recent LLMs
  • Finetune LLM using labelled examples.
  • Instruction Finetuning of LLMs.
  • Practical aspects of applications.

Materials The workshop will provide participants with all the materials they need to complete the exercises. These materials will include a workshop notebooks, datasets and codes.

Topics to be covered in the Workshop

  • State of Finetuning (Talk)
  • Inferencing: Prompting LLMs for a task
    • zero-shot
    • few-shot
  • Memory requirements (Talk)
  • Few Shot Finetuning using PEFT Techniques
  • Performance comparison of Prompting vs Few shot Finetuning on two domain specific datasets.
  • Building a Replication model (Talk)
  • Instruction Finetuning on Sequence to Sequence Task
  • Key Takeaways & Closing Talk


  1. Laptop with internet connection.
  2. Basic knowledge of using Python in Machine Learning.
  3. Understanding of Large Language Model.
  4. Familiarity with HuggingFace and Collab.

Content URLs:

All notebooks and slides can be download from GitHub link here.

Speaker Info:

I am an applied data scientist and research professional with 10+ years of relevant experience in solving problems leveraging advanced analytics, machine learning and deep learning techniques. I started my career as a Scientific Officer in a central government research organization (Bhabha Atomic Research Center) and worked on variety of domains such as conversational speech, satellite imagery and texts. Currently, I am working as a Director, Data Scientist with Fidelity investment for last 4 years working on language models (NLP) and Graphs.

As part of my work, I have used python throughout my career for solving data science problems as well as for pursuing research. I have published several academic and applied research papers and participated in multiple conferences over years. In past, I had trained professionals in machine learning and had been guest lecturer at BITS, Pilani, WILP program for Machine Learning subject (MTech course).

Speaker Links:




Open Source Contributions:

  1. finbert-embedding:
  2. classitransformers:
  3. PhraseExtraction

Section: Data Science, AI & ML
Type: Workshops
Target Audience: Intermediate
Last Updated: