AI final project --- STEAM (Steam top expert analysis machine)

• Llama 2 is licensed under the LLAMA 2 Community License, Copyright (c) Meta Platforms, Inc. All Rights Reserved.

• The project is a final project for NYCU CS AI course.

Introduction

• It's a project that can analysis the hot comments of the steam game.
• By clicking down a game name in UI, this program will automatically search the top comments of the game, and llama will organize top comments as a brief summerization.
• If you want to know whether a game is good or not, but you don't rely on the grading, besides, you don't want to look reviews one by one. This program is extraordinary suit for you.

On local machine

How to set up

• programming language
  • We are recommend to use python 3.10 or higher.
• packages
  • clone this project from GitHub, which means you run following command in your terminal, git clone https://github.com/Mushroom-MSL1L/NYCU-AI-Final-Project-STEAM.git
  • cd to this folder NYCU-AI-Final-Project-STEAM
  • run pip install -r requirements.txt
• model
  • This repo do not provide model, you need to download a one.
  • Make a folder whose name is llama.cpp under Model folder. And cd to llama.cpp folder.
  • Download llama-2-7B.Q4_k_M model as gguf file by running this command huggingface-cli download TheBloke/Llama-2-7b-Chat-GGUF llama-2-7b-chat.Q4_K_M.gguf --local-dir . --local-dir-use-symlinks False
  • Rename the file name as llama-2-7b-chat.Q4_K_M.gguf if needed.
  • There should has a path like NYCU-AI-Final-Project-STEAM\Model\llama.cpp\llama-2-7b-chat.Q4_K_M.gguf
  • More information can take a look at supplement.md.

How to use

• make sure your current folder is NYCU-AI-Final-Project-STEAM.
• run chainlit run ui.py -w in terminal
• open your browser and go to http://localhost:8000
• Type one game name you want to analysis
  • we recommend below games to test
    • "Cyberpunk 2077"
    • "Forza Horizon 4"
    • "Forza Horizon 5"
    • "The Witcher 3: Wild Hunt"
  • get more games from game_list.txt
• wait for a while, you will see the result.

On Colab (Recommend)

How to set up

• Upload STEAM.ipynb to google Colab notebook.
• Run the code one by one in the notebook.
  • Download llm model part may need 30 seconds
  • llama-cpp-python GPU ver. install part may need 20 minutes
• In Run part you need to update add-authtoken with your own token.
  • go to ngrok and sign up.
  • get your token from your-authtoken
  • Just press the copy button and paste it to the code.
  • like !ngrok config add-authtoken 2hYXWxSKTePaAIU6Xb86s9EJgYa_53qzsH6jSz88QHBaKt4Ad
• Keep run the remaining code, until it shows http://localhost:8000.

How to use

• Click the link http://localhost:8000 in the output.
• Type one game name you want to analysis
  • we recommend below games to test
    • "Cyberpunk 2077"
    • "Forza Horizon 4"
    • "Forza Horizon 5"
    • "The Witcher 3: Wild Hunt"
  • get more games from game_list.txt
• wait for a while, you will see the result.

Hyperparameter

model.py configuration

In class LLM, you are able to change some parameter in self.config.

• max_token: model terminates its output as it reaches the max token.
• n_ctx: context window, input length for the llm. Here are the parameters for both the GPU and the CPU.
• n_batch: batch size for the model. Here are the parameters for both the GPU and the CPU.
• n_gpu_layers: num of gpu_layers for the model. Available only for GPU.

chain.py configuration

In class chain in chain.py you are able to change some parameters in self.config.

• keywords:
  • keywords for retrieving. But, you have to concider the max token of input if you add more.
• total_document_length:
  • For max total_document tokens in prompt. total_document_length + template_prompt (approx. 40 words) cannot exceed n_ctx in Model/LLM
• max_docs_length:
  • For max document length for each document in prompt.max_docs_length * num_of_keywords must less than total_document_length at least.
• add_review_number:
  • the number of review you want to add in vectordatabse for each games.

Contributors ✨

Thanks goes to these wonderful people (emoji key):

Lu Junyou 💻 🐛 ✅ 👀

WilsonNYCU 📢 💻📹

Kennethii2i 💻 🚧

This project follows the all-contributors specification. Contributions of any kind welcome!

Access supplement.md to know more details