This is the codebase for UMD CMSC421 icl project. We provide methods to do in-context learning on two datasets---sst2 and gsm8k with Llama-2 chat version.
By default you only have 20G of disk space, which is definitely not enough for Llama-2-13B. Thus, you need to read the code first. Start with the main python files you run, i.e., the ones in the run folder. The variable prompt_dict is imported from batcon/consts.py. You can easily locate the other imported classes. Read these classes. All the namings are straightforward. After you read the code and have a plan about where to modify, let me know in the Discord channel. If you want to get a sense of what the code is doing, replace the default model with the 7B version (Read the code for how to do this. I haven't tested for it but I think the space should be roughly enough.) and follow the instructions below.
To use Llama-2, you need to first apply for access on the official website. After you submit the form, you will soon get an email from Meta titled "Get started with Llama 2". This means that you are granted access to Llama-2 for 24 hours. The next step is to fill in the application form on HuggingFace. I'm not sure if you can still do this after the 24 hours limit. Wait for some time and your HuggingFace account will be granted Llama-2 access. After that, find your HuggingFace access token in your account setting. Copy the token and paste it to configs/huggingface.token.
In the folder that contains this readme file, run
pip install -e .
To test the code on the first 2 examples of sst2, run
python run/few_shot_glue.py --valid_limit 2 --verbose
I suggest that you redirect the output to a text file like
python run/few_shot_glue.py --valid_limit 2 --verbose > output_sst2.out
To test the code on the first 2 examples of gsm8k, run
python run/few_shot_gsm.py --valid_limit 2 --verbose > output_gsm.out
Check the output files to get a sense of what the code is doing.