LLM4SC

Getting Started

Code Explanation

• answer: for each task, the classification results will be saved here

• config: for each task and model, the training configurations are saved here, including all four tasks and its configurations, all yaml files.

• data/dataset: datasets for all tasks

• exp_ML: experiments for Machine Learning Models.

• llm: old version for llms training(gpt2, qwen, llama)

• models: all models testing in the paper, including RNN-based(LSTM, GRU), CNN-based(three-layer CNN, ResNet), transformer(vanilla-transformer), llm(gpt2, qwen, llama)

• scripts: bash scripts for instructions

• utils: embedding layers

• few.py: Few-Shot tasks

• zero.py: Zero-Shot tasks

• run.py: Main experiment

• requirements.txt: all requires for training

Main Experiments

1. Install requirements. pip install -r requirements.txt

2. Download LLM Models. You can download all llms on Huggingface, and then save them in /llm/gpt2, /llm/Llama-3.2-1B, /llm/Qwen2.5-0.5B

3. Training. All the scripts are in the directory ./scripts. For example, if you want to get training for gpt2 on Actions-im dataset-1, just run the following command, and you can open ./logs/gpt2/action-im to see the results once the training is done:

   bash ./scripts/action-im.sh

Few-Shot and Zero-Shot Experiments

There are some examples in ./scripts. For example, if you want to get training for gpt2 on Phonation dataset-1 on Zero-Shot, just run the following command, and you can open ./logs/zero/pro1_zero_gpt2.log to see the results once the training is done:

bash ./scripts/pro1_zero.sh