plageon/SlimPlm

Small Models, Big Insights: Leveraging Slim Proxy Models To Decide When and What to Retrieve for LLMs

SlimPLM

📝 Paper • 🤗 Hugging Face • 🧩 Github

🌹 If you use this model, please star our GitHub repository to support us. Your star means a lot!

The code repository for ACL 2024 main conference paper: Small Models, Big Insights: Leveraging Slim Proxy Models To Decide When and What to Retrieve for LLMs.

⚡ News

May 2024: Our new work, Small Models, Big Insights: Leveraging Slim Proxy Models To Decide When and What to Retrieve for LLMs, has been accepted by ACL 2024 main conference.

Pre-requisites

Environment

Prepare python environment with the following command:

# https://github.com/castorini/pyserini/blob/master/docs/installation.md
conda create -n pyserini python=3.10 -y
conda activate pyserini
# install jdk and maven
conda install -c conda-forge openjdk=11 maven -y
# https://pytorch.org/get-started/locally/
# https://pytorch.org/get-started/previous-versions/
conda install pytorch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 pytorch-cuda=11.7 -c pytorch -c nvidia
# https://github.com/facebookresearch/faiss/blob/main/INSTALL.md
conda install -c conda-forge faiss-gpu
pip install pyserini pymongo django
# -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install scikit-learn transformers transformers[deepspeed] rouge_score evaluate dataset gpustat

Build a vllm environment for Qwen-72B-Chat inference with the following command:

# (Optional) Create a new conda environment.
conda create -n vllm python=3.9 -y
conda activate vllm

# Install vLLM with CUDA 12.1.
pip install vllm
pip install flash-attn --no-build-isolation

Search Corpus Preparation

Build kilt data corpus, please refer to kilt for more details.

# prepare mongodb for kilt data corpus
conda install -c anaconda mongodb
conda install -c conda-forge mongo-tools
wget http://dl.fbaipublicfiles.com/KILT/kilt_knowledgesource.json
mongoimport --db kilt --collection knowledgesource --file kilt_knowledgesource.json
mongod --dbpath ./mongodb/ --fork --logpath mongod.out

#rerank and retrieval API build up
#using django to build up the API
cd pyserini/searchServer
nohup python ./manage.py runserver 0.0.0.0:5050 > django.out &

Data Preparation

For ASQA, NaturalQuestions, TriviaQA, and ELI5, please download the data from kilt.

For MuSiQue, please download the data from MuSiQue.

The sampled test set ids are provided in user_intent_data directory.

SlimPLM

Model Fine-tuning

Prepare fine-tuning data for retrieval necessity judgment model, please refer to prepare_data.ipynb.

We release our fine-tuning data for retrieval necessity judgment model in v0104.

Prepare fine-tuning data for query rewriting model, please refer to gpt4_annotation.sh and seperate_claims.sh.

The fine-tuning data for query rewriting model is released in v0118.

The fine-tuned retrieval necessity judgment model is released in huggingface, and the query rewriting model is released in huggingface.

Model Inference

Run chat inference for proxy language model or large language model, please refer to chat_inference.sh.

Run inference on the fine-tuned retrieval necessity judgment model or query rewriting model, please refer to query_rewrite_inference.sh.

Reproduce Results

Follow these steps to reproduce the results in the paper.

1. Reproduce Direct Chat results, run chat inference for proxy language model or large language model. Make sure the "prompt_method" is "without_search".
2. Run inference on retrieval necessity judgment model, make sure the "rewrite_model" is "v0104".
3. Parse llama judge output strings. Please refer to scripts in prepare_data.ipynb.
4. Run inference on query rewriting model, make sure the "rewrite_model" is "v0108".
5. Parse llama question rewrite & separate claim results. Please refer to scripts in prepare_data.ipynb.
6. Run search inference using scripts in scripts/search_pipeline_apply.sh.
7. Run rerank inference using scripts in scripts/rerank_passage.sh.
8. Run evaluation on large language model using rag. Please make sure the "prompt_method" is "v0118{proxy_model_name}_rewrite_search", such as "v0118llama7b_rewrite_search".

Evaluation

Run evalation for short form answer dataset or long form answer dataset, please refer to evaluation_scripts.sh.

Baselines

Run SKR knn baseline, please prepare data using prepare_data.ipynb, and refer to skr_knn.sh.

Run Self-Eval baseline, please refer to self_eval.sh.

Run ITER-RETGEN baseline, please refer to itergen_chat.sh and itergen_search.sh.

Run Self-Ask baseline, please refer to selfask_chat.sh, selfask_reader.sh, and selfask_search.sh.

Citation

If you use the code in this repository, please cite the following paper:

@inproceedings{Tan2024SmallMB,
  title={Small Models, Big Insights: Leveraging Slim Proxy Models To Decide When and What to Retrieve for LLMs},
  author={Jiejun Tan and Zhicheng Dou and Yutao Zhu and Peidong Guo and Kun Fang and Ji-Rong Wen},
  year={2024},
  url={https://arxiv.org/abs/2402.12052}
}