Models section

ChatGPT prompt tuning script

You can run the following command to conduct the experiment for call chatgpt api to do the test. The code includes datasets for ReClor, LogiQA and LogiQA-V2. Here is an example to use ChatGPT to do testing. You can use your own openai api key to replace the OPENAI_API_KEY in openai.api_key = "OPENAI_API_KEY".

python run_chatgpt_prompt.py

You can follow this web link to create and use your own openai api key.

Alpaca prompt tuning script

You can run the following command to conduct the experiment on Alpaca to do the test. The code includes datasets for ReClor, LogiQA and LogiQA-V2. Here is an example to use alpaca to do testing.

python run_alpaca_prompt.py

Convert model to support transformers

Currently alpaca-7B are trained and located at /data/qbao775/Explanation-Generation/qiming_alpaca_7B under hgx2.sail.cloud.nesi.nz server. If you want to use Alpaca do conduct experiment, you may need to use hgx2.sail.cloud.nesi.nz server to do the experiment. LLaMA is located in /data/LLaMA under both hgx1.sail.cloud.nesi.nz and hgx2.sail.cloud.nesi.nz server. But if you want to load the LLaMA to do fine-tuning and testing, you need to convert the model into the format into huggingface version as the following shown.

## Convert the LLaMA-7B to LLaMA-7B huggingface model
python transformers/src/transformers/models/llama/convert_llama_weights_to_hf.py \
    --input_dir ../../LLaMA/7B \
    --model_size 7B \
    --output_dir llama_7B_hf

LLaMA-based model instruction fine-tunning

If you want to fine-tune LLaMA, please follow the following script. Here is an example to replicate Alpaca using LLaMA. We refer the code from here. You need to git clone the stanford_alpaca firstly as the following. The dataset alpaca_data.json is also been provided under the stanford_alpaca project.

git clone https://github.com/tatsu-lab/stanford_alpaca.git
cd stanford_alpaca

And then conduct the follwoing fine-tuning. Cause our requirements have included the packages that been used in stanford_alpaca, so you do not need to run the pip install -r requirements.txt again.

## Fine-tuning the LLaMA-7B and replicate the Alpaca model
CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node=4 --master_port=2024 train.py \
   --model_name_or_path llama_7B_hf/llama-7b \
   --data_path ./alpaca_data.json \
   --bf16 True \
   --output_dir qiming_alpaca \
   --num_train_epochs 3 \
   --per_device_train_batch_size 4 \
   --per_device_eval_batch_size 4 \
   --gradient_accumulation_steps 8 \
   --evaluation_strategy "no" \
   --save_strategy "steps" \
   --save_steps 2000 \
   --save_total_limit 1 \
   --learning_rate 2e-5 \
   --weight_decay 0. \
   --warmup_ratio 0.03 \
   --lr_scheduler_type "cosine" \
   --logging_steps 1 \
   --fsdp "full_shard auto_wrap" \
   --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
   --tf32 True

If you met some problems when you fine-tune LLaMA, you can have a look at this link which I summarized some of my experience.

Other encoder-based large language models from ReClor leaderboard training and evaluating scripts

You can run the following command to conduct the fine-tuning experiment on encoder-based large language models (MERIt, AMR-LE and LReasoner) to do the training. The code includes datasets for ReClor, LogiQA and LogiQA-V2. Here is an example do conduct fine-tuning using ReClor dataset.

sh ./scripts/run_LLM_reclor.sh

You can also change the argument for only conducting test like removing the following arguments and only keep the --do_test.

--do_train \
--evaluate_during_training \

fine-tuning/evaluation scripts for NLI and PARARULE-Plus tasks

You can run the following command to conduct the experiment on Alpaca to do the test. The code includes datasets from GLUE including MNLI, MRPC, QNLI, RTE, QQP, PARARULE-Plus-Depth-2, PARARULE-Plus-Depth-3, PARARULE-Plus-Depth-4, and PARARULE-Plus-Depth-5.

python run_glue_no_trainer.py \
  --model_name_or_path bert-base-cased \
  --task_name qbao775/PARARULE-Plus-Depth-2\
  --max_length 256 \
  --per_device_train_batch_size 32 \
  --learning_rate 2e-5 \
  --num_train_epochs 3 \
  --output_dir /tmp/PARARULE-Plus-Depth-2/