Fine-tuning large language models made easy with LoRA Cookbook #NLP

This content provides a detailed guide on fine-tuning a Phi-3-mini model for Python code generation using LoRA through the Hugging Face Hub.

The process starts with installing necessary Python libraries like bitsandbytes, transformers, peft, accelerate, datasets, trl, flash_attn, wandb, huggingface_hub, python-dotenv, absl-py, nltk, and rouge_score.

Next, the libraries are imported, and the connection to the Hugging Face Hub is established either through interactive login or using an API token. Global parameters for model configuration, dataset configuration, output model configuration, device configuration, LoRA configuration, and random seed are defined.

The dataset is loaded, formatted for ChatML, and split into training and test sets. The tokenizer and model are loaded, LoRA and training arguments are configured, and Weights & Biases is set up for experiment tracking.

The trainer is initialized, the model is trained and saved locally, the adapter is pushed to the Hugging Face Hub, and the final model is merged, saved, and pushed to the Hub.

For inference, the fine-tuned model and tokenizer are downloaded, the dataset is prepared, a text generation pipeline is set up, and inference is performed based on a user prompt.

Finally, the model is evaluated using the ROUGE metric by calculating ROUGE scores for model responses compared to reference outputs. Instructions are provided on how to download, prepare, infer, and evaluate the fine-tuned model using a structured approach and standard metrics.

Source link

Source link: https://medium.com/@mshojaei77/the-lora-cookbook-fine-tuning-large-language-models-for-everyone-55029a35a2eb?source=rss——large_language_models-5