This repo was created for an assignment in my NLP/LLM class. The goal was to take pre-trained models and finetune them with a dataset to adapt them to generate text/code based on an instruction. These were the models used:
- https://huggingface.co/meta-llama/Llama-2-7b-hf/tree/main
- https://huggingface.co/microsoft/phi-2/tree/main
- https://huggingface.co/mistralai/Mistral-7B-v0.1/tree/main
- A GPU with lots of RAM (40GB+)
- Python 3.8+
Required Libraries/Modules made be installed with this command:
pip install accelerate peft bitsandbytes transformers trl rouge bert_score evaluate
https://huggingface.co/datasets/vicgalle/alpaca-gpt4
There's a python file to finetune each pretrained model. Each program prints the finetuned model's first 20 responses to the terminal. Each program, after generating a finetuned model, runs 10 trials with the model, using different values of top_k, num_beams, and temperature for each trial. Each trial prints the model's responses for 20 inputs.
To use the llama2 model, you will first need to get and set your own huggingface auth token in finetune_llama2.py:
access_token = 'YOUR_HF_TOKEN'
Then you can run the following command:
python finetune_llama2.py
python finetune_phi2.py
python finetune_mistral.py
In the trials, there are 4 different values used for temperature, top_k, and num_beams:
temperature: 1.00, 0.66, 0.33, 0.00top_k: 8, 4, 2, 1num_beams: 8, 4, 2, 1
Here I observed when num_beams was set to 1, the other params had no visible effects:
Params BLEU Rouge BERT Human_Eval
Temperature: 1.00, Top_K: 1, Num_beam: 1 0.0557 0.5922 0.7330 0.7810
Temperature: 1.00, Top_K: 1, Num_beam: 2 0.0557 0.4887 0.7422 0.8470
Temperature: 1.00, Top_K: 1, Num_beam: 4 0.0568 0.5327 0.7411 0.8300
Temperature: 1.00, Top_K: 1, Num_beam: 8 0.0589 0.6632 0.7355 0.7450
Temperature: 1.00, Top_K: 2, Num_beam: 1 0.0557 0.5922 0.7330 0.8450
Temperature: 1.00, Top_K: 4, Num_beam: 1 0.0557 0.5922 0.7330 0.7790
Temperature: 1.00, Top_K: 8, Num_beam: 1 0.0557 0.5922 0.7330 0.7790
Temperature: 0.66, Top_K: 1, Num_beam: 1 0.0557 0.5922 0.7330 0.7790
Temperature: 0.33, Top_K: 1, Num_beam: 1 0.0557 0.5922 0.7330 0.7620
Temperature: 0.01, Top_K: 1, Num_beam: 1 0.0557 0.5922 0.7330 0.7790
Params BLEU Rouge BERT Human_Eval
Temperature: 1.00, Top_K: 8, Num_beam: 8 0.1718 0.6624 0.6898 0.6785
Temperature: 1.00, Top_K: 8, Num_beam: 4 0.1269 0.6305 0.7167 0.6125
Temperature: 1.00, Top_K: 8, Num_beam: 2 0.0792 0.5168 0.7098 0.7800
Temperature: 1.00, Top_K: 8, Num_beam: 1 0.0911 0.5378 0.7020 0.7790
Temperature: 1.00, Top_K: 4, Num_beam: 8 0.1718 0.6624 0.6898 0.7115
Temperature: 1.00, Top_K: 2, Num_beam: 8 0.1718 0.6624 0.6898 0.6615
Temperature: 1.00, Top_K: 1, Num_beam: 8 0.1718 0.6624 0.6898 0.6950
Temperature: 0.66, Top_K: 8, Num_beam: 8 0.1718 0.6624 0.6898 0.6950
Temperature: 0.33, Top_K: 8, Num_beam: 8 0.1718 0.6624 0.6898 0.7115
Temperature: 0.01, Top_K: 8, Num_beam: 8 0.1718 0.6624 0.6898 0.7280
Params BLEU Rouge BERT Human_Eval
Temperature: 1.00, Top_K: 8, Num_beam: 8 0.0562 0.6818 0.7401 0.7280
Temperature: 1.00, Top_K: 8, Num_beam: 4 0.0567 0.6366 0.7407 0.7790
Temperature: 1.00, Top_K: 8, Num_beam: 2 0.0699 0.6727 0.7384 0.7280
Temperature: 1.00, Top_K: 8, Num_beam: 1 0.0658 0.3748 0.7495 0.8300
Temperature: 1.00, Top_K: 4, Num_beam: 8 0.0579 0.7058 0.7361 0.7110
Temperature: 1.00, Top_K: 2, Num_beam: 8 0.0560 0.6698 0.7385 0.7450
Temperature: 1.00, Top_K: 1, Num_beam: 8 0.0560 0.6698 0.7385 0.7450
Temperature: 0.66, Top_K: 8, Num_beam: 8 0.0565 0.6678 0.7370 0.7450
Temperature: 0.33, Top_K: 8, Num_beam: 8 0.0571 0.6939 0.7372 0.7280
Temperature: 0.01, Top_K: 8, Num_beam: 8 0.0560 0.6698 0.7385 0.7450
Model Name BLEU Rouge BERT Human_Eval
llama-2 0.0560 0.6698 0.7385 0.7450
phi-2 0.1718 0.6624 0.6898 0.6950
mistral 0.0589 0.6632 0.7355 0.7450
The code wis partially based on this datacamp tutorial: https://www.datacamp.com/tutorial/fine-tuning-llama-2