README.md

LLM (Large Language Models) FineTuning Projects and notes on common practical techniques

Find me here..

• 🐦 TWITTER: https://twitter.com/rohanpaul_ai
• 🟠 YouTube: https://www.youtube.com/@RohanPaul-AI/featured
• 👨🏻‍💼 LINKEDIN: https://www.linkedin.com/in/rohan-paul-b27285129/
• ​👨‍🔧​ KAGGLE: https://www.kaggle.com/paulrohan2020

---

Fine-tuning LLM (and YouTube Video Explanations)

Notebook	🟠 YouTube Video
Finetune Llama-3-8B with unsloth 4bit quantized with ORPO
Llama-3 Finetuning on custom dataset with unsloth
CodeLLaMA-34B - Conversational Agent
Inference Yarn-Llama-2-13b-128k with KV Cache to answer quiz on very long textbook
Mistral 7B FineTuning with_PEFT and QLORA
Falcon finetuning on openassistant-guanaco
Fine Tuning Phi 1_5 with PEFT and QLoRA
Web scraping with Large Language Models (LLM)-AnthropicAI + LangChainAI

---

Fine-tuning LLM

Notebook	Colab
📌 Gemma_2b_finetuning_ORPO_full_precision
📌 Jamba_Finetuning_Colab-Pro
📌 Finetune codellama-34B with QLoRA
📌 Mixtral Chatbot with Gradio
📌 togetherai api to run Mixtral
📌 Integrating TogetherAI with LangChain 🦙
📌 Mistral-7B-Instruct_GPTQ - Finetune on finance-alpaca dataset 🦙
📌 Mistral 7b FineTuning with DPO Direct_Preference_Optimization
📌 Finetune llama_2_GPTQ
📌 TinyLlama with Unsloth and_RoPE_Scaling dolly-15 dataset
📌 Tinyllama fine-tuning with Taylor_Swift Song lyrics

---

LLM Techniques and utils - Explained

LLM Concepts
📌 DPO (Direct Preference Optimization) training and its datasets
📌 4-bit LLM Quantization with GPTQ
📌 Quantize with HF Transformers
📌 Understanding rank r in LoRA and related Matrix_Math
📌 Rotary Embeddings (RopE) is one of the Fundamental Building Blocks of LlaMA-2 Implementation
📌 Chat Templates in HuggingFace
📌 How is Mixtral 8x7B is a dense 47Bn param model
📌 The concept of validation log perplexity in LLM training - a note on fundamentals.
📌 Why we need to identify target_layers for LoRA/QLoRA
📌 Evaluate Token per sec
📌 traversing through nested attributes (or sub-modules) of a PyTorch module
📌 Implementation of Sparse Mixtures-of-Experts layer in PyTorch from Mistral Official Repo
📌 Util method to extract a specific token's representation from the last hidden states of a transformer model.
📌 Convert PyTorch model's parameters and tensors to half-precision floating-point format
📌 Quantizing 🤗 Transformers models with the GPTQ method
📌 Quantize Mixtral-8x7B so it can run in 24GB GPU
📌 What is GGML or GGUF in the world of Large Language Models ?

---

Other Smaller Language Models

• DeBERTa Fine Tuning for Amazon Review Dataset Pytorch

• FineTuning BERT for Multi-Class Classification on custom Dataset

• Document STRIDE when Tokenizing with HuggingFace Transformer for NLP Projects

• Fine-tuning of a PreTrained Transformer model - what really happens to the weights (parameters)

• Cerebras-GPT New Large Language Model Open Sourced with Apache 2.0 License

• Roberta-Large Named Entity Recognition on Kaggle NLP Competition with PyTorch

• Longformer end to end with Kaggle NLP competition

• Zero Shot Multilingual Sentiment Classification with PyTorch Lightning

• Fine Tuning Transformer (BERT) for Customer Review Prediction | NLP | HuggingFace

• Understanding BERT Embeddings and Tokenization | NLP | HuggingFace

• Topic Modeling with BERTopic | arxiv-abstract dataset

• Latent Dirichlet Allocation (LDA) for Topic Modeling

• Adding a custom task-specific Layer to a HuggingFace Pretrained Model

• Fine Tuning DistilBERT for Multiclass Text Classification

• Fine Tuning BERT for Named Entity Recognition (NER)

• Text Summarization by Fine Tuning Transformer Model | NLP

• Text Summarization with Transformer - BART + T5 + Pegasus

• Debarta-v3-large model fine tuning for Kaggle Competition Feedback-Prize | NLP

• Topic Modeling with BERT and Automatic Cluster Labeling

• Decoding strategies while generating text with GPT-2

• Fake News Classification with LSTM and Tensorflow

• FinBERT Sentiment Analysis for very Long Text (more than 512 Tokens) | PART 2

• FinBERT Sentiment Analysis for very Long Text Corpus (more than 512 Tokens) | PART-1

• Cosine Similarity between sentences with Transformers HuggingFace

• Zero Shot Learning - Cross Lingual Named Entity Recognition with XLM-Roberta

• BERT from Hugging Face - Few Baseline Application | NLP

• Transformer Encoder with Scaled Dot Product from Scratch

• Fuzzy String Matching in Natural Language Processing | NLP

• Understanding Word Vectors usage with Spacy Word and Sentence Similarity

• Named Entity Recognition NER using spaCy - Extracting Subject Verb Action

• Fine-Tuning-DistilBert - Hugging Face Transformer for Poem Sentiment Prediction | NLP

• Fine Tuning BERT-Based-Uncased Hugging Face Model on Kaggle Hate Speech Dataset

• Text Analytics of Tweet Emotion - EDA with Plotly

• Sentiment analysis using TextBlob and Vader