This project aims to provide some guidence on Fine tuning LLMs

Tools

• Jupyter Notebook
• Cohere: Provides some powerful LLMs, free and limited API requests for researchers. Cohere
• HuggingFace: A platform with huge amount of open-source LLMs, DataSets and so on. Huggingface
• Docanno: A tool for labeling the dataset and can be used to build instruction + output for training. Docanno

General Steps

1. Find the Data Source :You can take advantage of some open source data, or Crawling some data on the Internet (Please note the legality).

2. Pre-processing the dataset: Filter the data from the data source, leave only the information needed for further processing or training. Better to structure it to some structured format(e.g. json).

​

​ Use LLMs to help:

• pre-processing_data/MT_Filter_5TEs_Huggingface.ipynb : This file provides an example to use the model on huggingface to help with processing data.
• pre-processing_data/MT_Filter_ISO_Cohere.ipynb : This file provides an example to use the model provided by cohere to processing data.

3. Processing the dataset: Before training, you need to build you own dataset for training, typically including Instruction + Output. Docanno can be used as a tool.

• processing_data/convert_dataset_to_train.ipynb : This file provides an example of converting the dataset to the format of the model needed.
• processing_data/convert_jsonL.ipynb: This file provides an example of modifying dataset directly using json.
• processing_data/split_dataset_to_train_test.ipynb: This file provides an example of spliting the dataset into the training set and test set.

4. Training model（Fine Tuning）: After build your own dataset, you can start to fine tune a pre-trained model (choose an instruct pre-trained model from HuggingFace).

• training_model: This folder provides some examples to train the model using LoRA and upload the final model to HuggingFace.

5. Evaluation: After training the model, you can start evaluation, some metrics can be find in metrics:

• evaluation_scripts/predicting_with_models: This folder provides some examples for predicting with the final model.(Either use GPU or Inference API in HuggingFace)
• evaluation_scripts/evaluation_metric.ipynb: Gives an example of evaluation

Other scripts

• visualization.ipynb: Some template to Visualize data.
• CheckData.py: Check the data in the dataset.
• try_final_model.ipynb: Check if the trained model meets expectations.

Author: 0YHR0