Evaluation Framework

For a holistic evaluation, we will make use of the Evaluation Framework that contains 4 pillars:

• Performance
• Time to Train
• Cost to Train
• Inferencing

For each of the above four pillars, we are sharing our codebase and insights to:

• Assist you to leverage LLMs for your business needs and challenges
• Decide which LLM suits your needs from a performance and cost perspective
• Boost reproducibility efforts which are becoming increasingly difficult with LLMs

We are providing scripts that are ready-to-use for:

• Finetuning LLMs on your proprietary dataset via PeFT methodologies such as LoRA and Prefix Tuning
• Performing hyperparameter optimization to get the maximum performance out of these models

Getting Started

You can start fine-tuning your choice of LLM in 4 easy steps:

1. Setup conda environment

   wget https://repo.anaconda.com/miniconda/Miniconda3-py38_4.11.0-Linux-x86_64.sh
   bash Miniconda3-py38_4.11.0-Linux-x86_64.sh
   source ~/.bashrc
   conda create --name llm_finetuning python=3.9
   conda activate llm_finetuning
   

2. Install relevant packages

   git clone https://github.com/georgian-io/LLM-Finetuning-Hub.git
   cd LLM-Finetuning-Hub/
   pip install -r requirements.txt
   

3. Finetune your LLM of choice

   For instance, to finetune Llama2-7B or Llama2-13B, do the following:

   cd llama2/ # navigate to Llama2 folder
   python llama2_classification.py --lora_r 8 --epochs 5 --dropout 0.1 --pretrained_ckpt NousResearch/Llama-2-7b-hf # finetune Llama2-7B on newsgroup classification dataset
   python llama2_classification_inference.py --experiment <experiment folder> # evaluate finetuned Llama2 7B version
   python llama2_summarization.py --lora_r 8 --epochs 1 --dropout 0.1 --pretrained_ckpt NousResearch/Llama-2-13b-hf # finetune Llama2-13B on samsum chat dataset
   python llama2_summarization_inference.py --experiment <experiment folder> # evaluate finetuned Llama2 13B version
   

   For instance, to finetune Falcon-7B, do the following:

   cd falcon/ # navigate to Falcon folder
   python falcon_classification.py --lora_r 8 --epochs 5 --dropout 0.1 # finetune Falcon-7B on newsgroup classification dataset
   python falcon_classification_inference.py --experiment <experiment folder> # evaluate finetuned Falcon
   python falcon_summarization.py --lora_r 8 --epochs 1 --dropout 0.1 # finetune Falcon-7B on samsum chat dataset
   python falcon_summarization_inference.py --experiment <experiment folder> # evaluate finetuned Falcon
   

   For instance, to finetune Flan-T5-Large, do the following:

   cd flan-t5/ # navigate to Flan-T5 folder
   python flan_classification.py --peft_method prefix --prefix_tokens 20 --epochs 5 # finetune Flan-T5 on newsgroup dataset
   python flan_classification_inference.py --experiment <experiment folder> # evaluate finetuned Flan-T5
   python flan_summarization.py --peft_method lora --lora_r 8 --epochs 1 # finetune Flan-T5 on samsum chat dataset
   python flan_summarization_inference.py --experiment <experiment folder> # evalute finetuned Flan-T5
   

4. Zero-shot and Few-shot your LLM of choice

   For instance, to use Falcon-7B on newsgroup classification task, do the following:

   python falcon_baseline_inference.py --task_type classification --prompt_type zero-shot
   python falcon_baseline_inference.py --task_type classification --prompt_type few-shot
   

   To use Falcon-7B on samsum summarization task, do the following:

   python falcon_baseline_inference.py --task_type summarization --prompt_type zero-shot
   python falcon_baseline_inference.py --task_type summarization --prompt_type few-shot
   

NOTE: All of our experiments were conducted on the AWS EC2 instance: g5.2xlarge. It has one 24GB Nvidia GPU, and is sufficient to finetune the LLMs in this repository.

LLM Roadmap

Our plan is to perform these experiments on all the LLMs below. To that end, this is a tentative roadmap of the LLMs that we aim to cover, and their corresponding codebase and README links:

Benchmarks

We benchmark LLMs across the tasks of classification and summarization. More precisely, we finetune LLMs on the following experiments:

1. Classification: Sample efficiency VS Accuracy
2. Summarization: Fine-tuning

Note:

• RP refers to RedPajama.

Classification: Sample efficiency VS Accuracy

We use the Newsgroup dataset which is a 20-way classification problem. Each document needs to be identified as one of the 20 possible newsgroups. To check how quickly LLMs can learn on small number of samples, we compare them with the likes of BERT and Distilbert. Following table captures how models perform as we increase the number of training samples.

Summarization: Finetuning

We use the samsum dataset which contains chat conversations and their summarized versions. The task here is for LLMs to learn how best to summarize conversations by learning from pairs of conversations and corresponding summaries. Following table captures how LLMs perform on this task.

Contributing

If you would like to contribute to this project, we recommend following the "fork-and-pull" Git workflow.

1. Fork the repo on GitHub
2. Clone the project to your own machine
3. Commit changes to your own branch
4. Push your work back up to your fork
5. Submit a Pull request so that we can review your changes

NOTE: Be sure to merge the latest from "upstream" before making a pull request!

Correspondence

If you have any questions, please reach out to:

• Rohit Saha (Email | LinkedIn)
• Kyryl Truskovskyi (Email | LinkedIn)
• Maria Ponomarenko (Email | LinkedIn)