LLMs Engineering : From Using LLMs to Training LLMs

Explore how APIs connect apps to closed and open source LLMs, and how tokens drive cost through tokenization, comparing OpenAI's ChatGPT, Gemini, DeepSeek.

Set up a Jupyter notebook, load your secret key from secretkey.py, create an OpenAI client, and request a response with GPT-5 nano; inspect the output text and id.

Demonstrate how to generate and compare multiple responses to one question by varying temperature settings, looping prompts, and printing results, using a nano model and code two.

Learn to build a conversation loop for a chatbot, manage messages with roles and user inputs, and guide an assistant to act as a sales voice while handling exit conditions.

Explore prompt engineering by building a first application with the OpenAI API, using secret keys, and pandas to create a student Q&A data frame for scoring.

Discover prompt engineering to grade student answers with an LLM by constructing a dynamic prompt, applying historical accuracy and completeness rubrics, and returning structured JSON grades.

Learn how to access APIs for solvers like DeepSeq, obtain API keys, manage paid credits, and implement secret keys in Python to create your own DeepSeek solutions.

Explore retrieval augmented generation (RAC) and how to build it from document collection and text extraction to chunking, metadata, vector storage, and context-driven answer generation.

Build a chatbot using a rag model by embedding questions into vectors, selecting the top five chunks as context for an LLM to answer only from the provided text.

Configure multi-source data by creating a data folder with country metadata. Extract text from PDFs into documents, chunk paragraphs, and save embeddings and metadata.

Explore huggingface, the open-source platform for models, datasets, and transformers that powers LLMs with translation, summarization, QA, classification, and multi-modal supports.

Install transformers with pip, choose CPU or GPU via torch, and run a sentiment analysis pipeline using a Hugging Face model to test open source capabilities.

Explore how to use HuggingFace models to build task-oriented solutions with zero-shot classification, leveraging the Transformers pipeline to classify text into custom labels.

Leverage the huggingface transformers pipeline for image object detection, processing a test image with pillow, returning labeled detections and pixel coordinates, then extend to question answering and text summarization.

Learn to build a text summarizer with a transformer-based sequence-to-sequence model by loading a pre-trained auto model and tokenizer to generate concise summaries.

Build a chat bot by importing transformers and using a pipeline with a small model, while managing conversation history and generation settings like temperature.

Load local data with embeddings, enable dangerous deserialization, retrieve a limited number of chunks, and run an LLM with a context prompt to answer tax filing questions.

Prepare dictionary-style text data in the required format for the tokenizer and model using the dataset library, and apply a train-test split of 0.2 before fine-tuning.

Train a model quickly using Hugging Face transformers with tokenization and sequence classification. Build the model, tokenize the data, set training arguments, and evaluate accuracy before and after training.

Save and load a trained model and tokenizer, build a classification pipeline, and evaluate performance for predictions; retrain to improve accuracy.

Fine-tune a model by adjusting training arguments and epochs, save the new model, load it for evaluation, and track accuracy gains with loss reduction.

Explore PyTorch tensors from scalars to vectors, matrices, and higher-dimensional arrays, and learn to create, convert, index, and reshape with random values and data types.

Explore connecting tensors with NumPy by converting between PyTorch tensors and NumPy arrays, including handling GPU tensors and ensuring CPU conversion for NumPy operations.

Learn how backpropagation trains deep neural networks by adjusting weights and biases through forward and backward passes, applying the chain rule to minimize mean square error.

Explore how the transformer neural network powers LLMs, detailing encoder-decoder architecture, embeddings, positional encoding, and multi-head attention with QKV to predict the next word or character.

Download Hamlet text from archive.org, save it as a text file, and load it into a Jupyter notebook with utf-8 encoding to prepare a Shakespeare-like model.

Learn to encode text into numbers to create data using torch.tensor and torch.long, build 1d tensors, and split data into 90% training and 10% validation to train a model.

Set up a transformer model using open-source code, configure patch size, embeddings, heads, layers, dropout, then train with Adam on GPU or CPU and generate a text of 500 characters.

Explore how model training transforms outputs from random gibberish to coherent text, showing the impact of iterations, evaluation steps, and larger training runs across languages.