https://github.com/kwokhing/ai-planet-llm-bootcamp-challenge

An LLM challenge to (i) fine-tune pre-trained HuggingFace transformer model to build a Code Generation language model, and (ii) build a retrieval-augmented generation (RAG) application using LangChain
https://github.com/kwokhing/ai-planet-llm-bootcamp-challenge

embeddings-model fine-tuning langchain language-model llm mistral-7b ocra-mini-3b qlora retrieval-augmented-generation sentence-embeddings supervised-finetuning transformer-models

Last synced: 4 months ago
JSON representation

An LLM challenge to (i) fine-tune pre-trained HuggingFace transformer model to build a Code Generation language model, and (ii) build a retrieval-augmented generation (RAG) application using LangChain

• Host: GitHub
• URL: https://github.com/kwokhing/ai-planet-llm-bootcamp-challenge
• Owner: KwokHing
• Created: 2023-11-16T01:25:08.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2023-11-23T02:58:53.000Z (over 1 year ago)
• Last Synced: 2024-12-02T20:49:22.742Z (6 months ago)
• Topics: embeddings-model, fine-tuning, langchain, language-model, llm, mistral-7b, ocra-mini-3b, qlora, retrieval-augmented-generation, sentence-embeddings, supervised-finetuning, transformer-models
• Language: Jupyter Notebook
• Homepage: https://aiplanet.com/bootcamps/llm-bootcamp
• Size: 874 KB
• Stars: 4
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
## AI Planet Large Language Model Bootcamp Challenge

A challenge from AI Planet's LLM bootcamp to (i) fine-tune pre-trained HuggingFace transformer model to build a Code Generation language model, and (ii) build a retrieval-augmented generation (RAG) application using LangChain

---

#### Part I: Fine-tuning [Orca Mini 3B](https://huggingface.co/pankajmathur/orca_mini_3b) on [evolved codealpaca](https://huggingface.co/datasets/theblackcat102/evol-codealpaca-v1) dataset to build a Code Generation model

Fine-tuning Orca Mini 3B with the evolved codealpaca dataset equips Orca Mini the ability to perform better on code generation tasks. 

When we asked Orca Mini to provide us codes for scikit-learn linear regression before fine-tuning was done, it requested for more information to be provided. 

![jpg](img/orca-3b-before-finetuned.png)

After fine-tuning, the model is able to return a more targeted code generation completion output shown below.

![jpg](img/orca-3b-after-finetuned.png)

---

#### Part II: Building a Question & Answering Retrieval-Augmented Generation (RAG) application using LangChain

A simple RAG to answer questions on the Battle Line game rules (PDF) using LangChain. We used the INSTRUCTOR embeddings model (ranked highly on the Massive Text Embeddings Benchmark, [MTEB](https://huggingface.co/spaces/mteb/leaderboard) leaderboard) for performing semantic retrieval and a quantized version of the impressive [Mistral 7B](https://mistral.ai/product/) for returning the completion output. 

Examples of the RAG output:

![jpg](img/Simple-Mistral-RAG-example.png)

![jpg](img/Mistral-RAG-example.png)

---

## Getting started

Open `AI_Planet_LLM_Bootcamp_Challenge.ipynb` on a jupyter notebook environment. Alternatively, you can view the codes in [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1TU4XPGS04xdMHY1wXj23Z9OiELp7XJ_7#scrollTo=-gZxQYC5g4No?usp=sharing). The notebook consists of further technical details.

---

## Bonus - Example on Fine-tuning Phi 1.5B using QloRA

Open `AI_Planet_Bootcamp_Final_Assignment_Fine_tuning_Phi_1_5b.ipynb` on a jupyter notebook environment. The notebook consists of further technical details.