First, create a conda environment named RAG4IE and install the required packages:
conda create -n RAG4IE python=3.10
conda activate RAG4IE
pip install -r requirements.txtNote: TACRED is licensed by the Linguistic Data Consortium (LDC), so we cannot directly publish the prompts or the raw results from the experiments conducted with Llama and Mistral, since the responses of these models consist of the prompts in their instruction parts. However, we have published the returned results when Llama and Mistral were integrated. Upon an official request, the data can be accessed on LDC, and the experiments can be easily replicated by following the instructions provided.
Change the paths and configurations under config.ini for your experiment.
Place the following datasets under the data folder:
-
TACRED: Licensed by the Linguistic Data Consortium (LDC). Please download it from here.
-
WikiEvents: Available on Github and place it under the
datafolder. -
RAMS: Available on Github and place it under the
datafolder. -
ACE2005: Licensed by the Linguistic Data Consortium (LDC).
-
SemEval: Available on Hugging Face and place it under the
datafolder.
pip install -r requirements.txtCompute embeddings and similarities for benchmark datasets in advance:
cd src/data_augmentation/embeddings
python sentence_embeddings.py
python sentence_sim.pyReplace all specific paths with replace_with_your_path:
python main.py \
--test_data_path replace_with_your_path/data/rams/test.jsonlines \
--train_data_path replace_with_your_path/data/rams/train.jsonlines \
--similar_sentences_path replace_with_your_path/data/rams/test_rams_similarities.json \
--dataset rams \
--prompt_type rag \
--model_name gpt3.5 \
--responses_path replace_with_your_path/outputs/rams/gpt3.5_rams_rag_5_doc.json \
--topk 5 \
--task EEpython main.py
--test_data_path replace_with_your_path \
--train_data_path replace_with_your_path \
--similar_sentences_path replace_with_your_path \
--dataset semeval \
--prompt_type rag \
--model_name LLaMA3-8b-instruct \
--responses_path replace_with_your_path \
--topk 10 \
--task RESimilarly, replace specific paths with replace_with_your_path:
accelerate launch \
--config_file replace_with_your_path/src/train/configs/sft_zero1.yaml \
--num_processes 4 \
--num_machines 1 \
--machine_rank 0 \
--deepspeed_multinode_launcher standard \
replace_with_your_path/src/train/train_llm.py \
--experiment_name EAE_train_no_retrieval \
--model_path replace_with_your_path/models/Meta-Llama-3-8B-Instruct/ \
--max_ckpts 3 \
--max_seq_len 2048 \
--gradient_accumulation_steps 16 \
--output_dir ./ckpts \
--log_dir ./train_logs \
--n_epochs 3 \
--train_bsz_per_gpu 2 \
--eval_bsz_per_gpu 2 \
--learning_rate 2e-5 \
--eval_step 100 \
--save_step 100 \
--gradient_checkpointing