daywatch/LLM_and_ChatGPT

LLM and ChatGPT work in progress!

Skills and projects covered in LLM and ChatGPT folders:

Section 1: Large Language Models: Application through Production (EdX)

LM1

Use Hugging Face datasets and large models; set up tokenizer config

LM2

Build a knowledge-based question answering / search system

• Convert a dataset to vectors and save them in a vector library (FAISS) or database (Chroma)
• Vectorize a query (with filters in it) and saved the output as context
• Combine context with original prompt as new prompts to generate search results

This module also has tutorials on Pinecone and Weaviate

LM3

Build tree LLM-chain-based models

• LLM1 moderates the comments generated by LLM2
• Use LLM with LangChain agents (Wiki, Google, Python REPL) to do automatic data analyses
• An LLM agent that allows user to have free chat with documents (e.g., Shakespeare's books)

LM4

Fine-tune LLMs with Hugging Face, Tensorboard, and DeepSpeed (multiple GPU cluster support) on a traditional IMDB classification; evaluate summarization performance with NLTK and ROUGE

LM5

Hugging Face Disaggregator (for q quick demographic analysis) and evaluate (for toxicity), gender expression generation, and SHAP (for interpretability, i.e., token-level contribution for the final generated output)

LM6

MLOps of a sample model with MLFlow, focusing on model registeration, versioning, monitoring/performance Tracking, and pushing to the production

Section 2: my fine-tuning practices:

• use QLoRa and hf PEFT to tune GPT-neo model on casual language modeling
• use deepspeed to tune mpt model on instruction data
• use XTurning on LLaMA 7b on question answering

Section 3: Generative AI and Large Language Models (Coursera)

Lab1: prompt-engineering

Lab2: fine-tune with LoRA

Lab3: reinforcement learning with human feedback (RLHF-PPO & KL)

Section 4: openai_api_ and prompt dev

• prompt engineering demos and guideline
• use of langchain, including prompt management, external agents, and evaluation tools
• fine-tuning service on chatgpt with generated data from gpt4
• A tutotial of AutoGen, a multi-agent LLM package

Section 5: Foundation Models from the Ground Up (EdX)

• transformer algorithm in torch
• fine-tuning with LORA and PEFT
• fine-tuning with prompt-tuning and PEFT
• quantization and deploylment

Section 6: Apply LLMOps using VetexAI(DeepLearningAI - GCP)

course notes

• L2: data ingestion using SQL and storage on GCP
• L3: use kubeflow for automation orchestratian + fine-tuning on vetexAI with QLoRA
• L4: make predictions and pull out metrics on safety scores, such as harassment and severity

Section 7: Azure AI APIs

This section includes several cases of Azure AI services, including Azure OpenAI, AI Studio, Speech (TTS & STT), AI Search, Document Intelligence, Storage, Vision, Language Service (translation, sentiment analysis, lang detection, intent & entity), and Semantic Kernel (similar to LangChain)

Section 8: Evaluate with Langfuse

Iterative development is crucial for LLM apps. I use a platform called Langfuse, similar to W&B and MLFlow. The code includes setting up Llama3 on Groq, logs, prompts, and datasets for LLM monitoring and improvement.

other resources

• A sample of my reserach links on LLM topics working as an Applied Scientist (NLP), including latest model update, prompt engineering, practice on fine-tuning and alignment, literatures, and downstream tasks and so on
• Fine-tuning huggingface LLMs on AWS SageMaker
• Fine_tuning huggingface LLMs within azureml