The process for fine-tuning LLMs Locally without third-party APIs and platform fees. Before starting to fine-tune a model, make sure you actually need to! Fine-tuning LLMs can be a lengthy process and computationally expensive. It is a good idea to test the model's performance on your task-specific data and try other less-demanding methods such as In-Context-Learning (ICL). It is also a good practice to fine-tune with smaller batches of data and record the improvement (i.e., output accuracy and validation loss) at each step, before increasing the number of training data examples.

-- Install dependencies as in dependencies.py for the sentiment analysis task.

-- Select a pre-trained model with fewer parameters, preferably a lightweight one that is easier to load and train.

-- Model specifications including model name, tokeniser, and sequence classifier (for classification tasks), as in model-spec.

-- Set the parameters for the main task function, e.g., sentiment analysis as in function(e.g., padding, truncation, prediction labels, etc). Specify the number of examples to provide to the LLM from the external data source; the dataset format should follow the LLM's training template.

-- Load the dataset you want to fine-tune your model on (its format should match the LLM's training template); you can use Pandas DataFrame to sort/select a portion of the training dataset, as in load_data.

-- Tokenisation: tokenise the training and evaluation sets of the data (specify the number of examples). Similar to model spec except we use the dataset for tokenisation as in specify_data.

-- Initialise the model with the 'function` as before with tokenised training and evaluation sets.

-- Training or fine-tuning the model using the Trainer module from the Transformers library as in training. Check the extended set of arguments at the Trainer class of TRansformers

This video tutorial is the easiest way to fine-tune a large language model 100% free and locally without using any API or third-party platform.

Terms and concepts discussed: Bidirectional Encoder Representations from Transformers, tokenisation, embedding, encoding, task head, encoder-only Transformer models, sequence classification