🎓 Google for Developers: Machine Learning Bootcamp Korea 2024

---

Under Construction 🚧

Starting from 20205.07.28 and onwards,this repository is currently under active development.

Please check back later for updates!

---

Acknowledgments:

🏅 Special thanks to the bootcamp organizers for their support and mentorship throughout this project journey!
- Kihun (Sam) Choi

📑 For detailed project documentation, refer to the docs/ folder!

📜 Historical Reference
For earlier versions of the project and documentation, refer to the OLD_README.md.

---

📑 Table of Contents

1. Overview
2. Enterprise Use Cases
3. About the Dataset
4. Model Architecture
5. Training and Checkpoints
6. Installation and Quickstart
7. Sample Output
8. ⚙️ Mac ARM64 Optimization Efforts
9. Limitations and Future Work

---

📌 1. Overview

A Secure, multi-LLM Enterprise Workflow/Productivity Empowering A.I. Solution

Key Capabilities:

Important

This project aims to build a sophisticated multi-LLM solution tailored for enterprise consulting use-cases, providing sentiment-aware, contextually enriched responses. The solution supports consultants in capturing user sentiment, detecting sarcasm, and generating impactful responses, enhancing client engagement and customer satisfaction.

• Customer Sentiment Analysis: Classifies feedback as positive, negative, or sarcastic to prioritize responses.
• B2B Support: Automates responses, saving time while maintaining professionalism.
• HR Feedback Analysis: Detects trends in employee feedback, identifying areas for improvement.

---

🏢 2. Enterprise Use Cases

1. Customer Support for SaaS and B2B Companies
   In fast-paced environments, Gemma-Sprint-Project flags high-risk tickets (negative or sarcastic tone) for prioritized handling, generating empathetic responses that increase customer satisfaction.

2. Real-Time Sentiment Analysis for Product Consultants
   This tool enables consultants to gain insights into customer sentiment across social media and reviews, identifying positive, negative, or sarcastic tones to support product improvement efforts.

---

📦 3. About the Dataset

Primary Dataset: UScomments_final_cleaned.csv

• Core Columns:
  • cleaned_comment: User comments, cleaned and pre-processed.
  • sentiment: Binary sentiment labels (1: Positive, 0: Negative).

Supplementary IT Glossary Dataset: IT_glossory_final_v4.csv

• Purpose: Provides a glossary of IT terms to enhance technical accuracy and relevance in responses, particularly for enterprise and IT consultancy.

---

⚙️ 4. Model Architecture

Model Stack:

1. GPT-J: Generates empathetic, context-rich responses.
2. MiniLM: Recognizes domain-specific IT terms efficiently.
3. DistilBERT: Detects sentiment for classification.
4. Sarcasm Detection Model: Identifies sarcasm for nuanced communication.
5. GEMMA-2B-IT: Delivers sentiment-aware, industry-specific responses.

---

📋 5. Training and Checkpoints

• Checkpoints: Stored in the checkpoints/ directory, minimizing data loss during lengthy computations.
• Logging: Real-time logging in logs/ allows progress tracking and troubleshooting.

Optimized for Apple Silicon (M2) using the MPS backend, balancing high-performance with efficient resource usage.

---

🚀 6. Installation and Quickstart

Pre-requisites:

1. Docker and Colima for containerized setup.
2. Python 3.10 with dependencies in requirements.txt.

Installation Steps

1. Clone the Repository:

   git clone https://github.com/dockerdoge12/gemma-sprint-prj.git

   cd gemma-sprint-prj

2. Pull Docker Image:

   docker pull dockerdoge12/gemma-sprint-prj:latest

3. Run Docker Container:

   docker run -p 8080:8080 dockerdoge12/gemma-sprint-prj:latest

---

📊 7. Sample Output

Loading checkpoint shards: 100%|████████████████████████████████████████████████████████████| 2/2 [00:01<00:00,  1.38it/s]

Sample Query 1:
User Comment: "We’re using legacy Oracle databases. How challenging would it be to migrate to a modern data warehouse solution?"

Response: Thank you for bringing this to our attention. Let's work through the compatibility concerns step by step.
**Understanding the Challenges**
- **Different Database Technologies:** Oracle Database is a relational database management system (RDBMS) based on the SQL standard. Fundamental difference in architecture can lead to compatibility issues.
- **Data Mapping and Transformation:**
   * Identify the data elements that need to be transferred between the two systems (e.g., customer master data, sales orders, inventory levels).
   * Determine the mapping between data types, structures, and formats in both databases. Specialized tools or custom scripts may be needed for data conversion.
- Would you like me to elaborate further?

---

⚙️ 8. Mac ARM64 Optimization Efforts

Optimizations were implemented to ensure model performance on Apple Silicon:

1. PyTorch over TensorFlow: Enabled compatibility with MPS (Metal Performance Shaders) for faster computation.
2. Batch Size Adjustment: Fine-tuned batch sizes to balance memory and computation on the M2 chip.
3. Reduced Computation Complexity: Limited training epochs and iteratively adjusted parameters to minimize resource usage.
4. Efficient Logging and Checkpointing: Detailed logs and checkpoints track performance without requiring re-runs.

---

🎯 9. Limitations and Future Work

Limitations

• Bias and Context Sensitivity: Sarcasm detection accuracy remains affected by contextual subtleties; further retraining with specific data is needed to enhance precision.
• Computational Constraints: Optimized for local M2 hardware, limiting training complexity and token length for efficient operation.
• Incomplete Responses: Implemented continuation prompts to address response truncation due to token limitations.

Future Work

1. Model Enhancements: Refine GEMMA-2B-IT prompts for better contextual accuracy.
2. Interactive UI Development: Build a real-time interface for improved user engagement and feedback integration.

---

Acknowledgments-GEMMA-Sprint Project:

1. DistilBERT for Sentiment Analysis Model: distilbert-base-uncased-finetuned-sst-2-english

   Source: "https://huggingface.co/distilbert-base-uncased-finetuned-sst-2-english"

   • Developed by Hugging Face, this model is fine-tuned for sentiment analysis and provides a compact version of BERT for efficient classification tasks.

2. Sarcasm Detection Model Model: helinivan/english-sarcasm-detector

   Source: Hugging Face Link: "https://huggingface.co/helinivan/english-sarcasm-detector"

   • Created by Helini Ivan on Hugging Face, this model is designed to detect sarcasm in English text, helping to capture nuances in sentiment.

3. GPT-J for Contextual Response Generation Model: EleutherAI/gpt-j-6B

   Source: Hugging Face Link: "https://huggingface.co/EleutherAI/gpt-j-6B"

   • Developed by EleutherAI, GPT-J is a powerful language model suitable for generating context-aware, coherent responses.

4. MiniLM for IT Domain-Specific Recognition Model: microsoft/MiniLM-L12-H384-uncased

   Source: Hugging Face Link: "https://huggingface.co/microsoft/MiniLM-L12-H384-uncased"

   • Created by Microsoft, this model provides efficient natural language understanding, especially for identifying and handling IT-specific terms in context.

5. GEMMA-2B-IT

   Source: Hugging Face

---