Awesome

RankLLM

We offer a suite of rerankers - pointwise models like monoT5 and listwise models with a focus on open source LLMs compatible with FastChat (e.g., Vicuna, Zephyr, etc.), vLLM or SGLang. We also support RankGPT variants, which are proprietary listwise rerankers. Some of the code in this repository is borrowed from RankGPT, PyGaggle, and LiT5!

Releases

current_version = 0.20.2

Note for Mac Users: RankLLM is not compatible with Apple Silicon (M1/M2) chips. However, you can still run it by using the Intel-based version of Anaconda and launching your terminal through Rosetta 2.

📟 Instructions

Create Conda Environment

conda create -n rankllm python=3.10
conda activate rankllm

Install Pytorch with CUDA (Windows/Linux)

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

Install Pytorch with MPS (Mac)

pip3 install torch torchvision torchaudio

Install openjdk with maven if you want to use the retriever

conda install -c conda-forge openjdk=21 maven -y

Install Dependencies

pip install -r requirements.txt

If building nmslib failed during installation, try manually installing the library with conda install -c conda-forge nmslib and following it up with pip install -r requirements.txt again.

Install vLLM or SGLang (Optional)

vLLM

pip install rank-llm[vllm]  # pip installation
pip install -e .[vllm]      # or local installation for development

SGLang

pip install rank-llm[sglang]  # pip installation
pip install -e .[sglang]      # or local installation for development

Remember to install flashinfer to use SGLang backend.

pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.4/

Run end to end - RankZephyr

We can run the RankZephyr model with the following command:

python src/rank_llm/scripts/run_rank_llm.py  --model_path=castorini/rank_zephyr_7b_v1_full --top_k_candidates=100 --dataset=dl20 \
--retrieval_method=SPLADE++_EnsembleDistil_ONNX --prompt_mode=rank_GPT  --context_size=4096 --variable_passages

Including the --vllm_batched flag will allow you to run the model in batched mode using the vLLM library.

Including the --sglang_batched flag will allow you to run the model in batched mode using the SGLang library.

If you want to run multiple passes of the model, you can use the --num_passes flag.

Run end to end - RankGPT4-o

We can run the RankGPT4-o model with the following command:

python src/rank_llm/scripts/run_rank_llm.py  --model_path=gpt-4o --top_k_candidates=100 --dataset=dl20 \
  --retrieval_method=bm25 --prompt_mode=rank_GPT_APEER  --context_size=4096 --use_azure_openai

Note that the --prompt_mode is set to rank_GPT_APEER to use the LLM refined prompt from APEER. This can be changed to rank_GPT to use the original prompt.

Run end to end - LiT5

We can run the LiT5-Distill V2 model (which could rerank 100 documents in a single pass) with the following command:

python src/rank_llm/scripts/run_rank_llm.py  --model_path=castorini/LiT5-Distill-large-v2 --top_k_candidates=100 --dataset=dl19 \
    --retrieval_method=bm25 --prompt_mode=LiT5  --context_size=150 --vllm_batched --batch_size=4 \
    --variable_passages --window_size=100

We can run the LiT5-Distill original model (which works with a window size of 20) with the following command:

python src/rank_llm/scripts/run_rank_llm.py  --model_path=castorini/LiT5-Distill-large --top_k_candidates=100 --dataset=dl19 \
    --retrieval_method=bm25 --prompt_mode=LiT5  --context_size=150 --vllm_batched --batch_size=32 \
    --variable_passages

We can run the LiT5-Score model with the following command:

python src/rank_llm/scripts/run_rank_llm.py  --model_path=castorini/LiT5-Score-large --top_k_candidates=100 --dataset=dl19 \
    --retrieval_method=bm25 --prompt_mode=LiT5 --context_size=150 --vllm_batched --batch_size=8 \
    --window_size=100 --variable_passages

Run end to end - monoT5

The following runs the 3B variant of monoT5 trained for 10K steps:

python src/rank_llm/scripts/run_rank_llm.py --model_path=castorini/monot5-3b-msmarco-10k --top_k_candidates=1000 --dataset=dl19 \
  --retrieval_method=bm25 --prompt_mode=monot5 --context_size=512

Note that we usually rerank 1K candidates with monoT5.

If you would like to contribute to the project, please refer to the contribution guidelines.

🦙🐧 Model Zoo

The following is a table of the listwise models our repository was primarily built to handle (with the models hosted on HuggingFace):

vLLM and SGLang backends are only supported for RankZephyr and RankVicuna models.

We also officially support the following rerankers built by our group:

LiT5 Suite

The following is a table specifically for our LiT5 suite of models hosted on HuggingFace:

Now you can run top-100 reranking with the v2 model in a single pass while maintaining efficiency!

monoT5 Suite - Pointwise Rerankers

The following is a table specifically for our monoT5 suite of models hosted on HuggingFace:

We recommend the Med models for biomedical retrieval. We also provide both 10K (generally better OOD effectiveness) and 100K checkpoints (better in-domain).

✨ References

If you use RankLLM, please cite the following relevant papers:

[2309.15088] RankVicuna: Zero-Shot Listwise Document Reranking with Open-Source Large Language Models

@ARTICLE{pradeep2023rankvicuna,
  title   = {{RankVicuna}: Zero-Shot Listwise Document Reranking with Open-Source Large Language Models},
  author  = {Ronak Pradeep and Sahel Sharifymoghaddam and Jimmy Lin},
  year    = {2023},
  journal = {arXiv:2309.15088}
}

[2312.02724] RankZephyr: Effective and Robust Zero-Shot Listwise Reranking is a Breeze!

@ARTICLE{pradeep2023rankzephyr,
  title   = {{RankZephyr}: Effective and Robust Zero-Shot Listwise Reranking is a Breeze!},
  author  = {Ronak Pradeep and Sahel Sharifymoghaddam and Jimmy Lin},
  year    = {2023},
  journal = {arXiv:2312.02724}
}

If you use one of the LiT5 models please cite the following relevant paper:

[2312.16098] Scaling Down, LiTting Up: Efficient Zero-Shot Listwise Reranking with Seq2seq Encoder-Decoder Models

@ARTICLE{tamber2023scaling,
  title   = {Scaling Down, LiTting Up: Efficient Zero-Shot Listwise Reranking with Seq2seq Encoder-Decoder Models},
  author  = {Manveer Singh Tamber and Ronak Pradeep and Jimmy Lin},
  year    = {2023},
  journal = {arXiv:2312.16098}
}

If you use one of the monoT5 models please cite the following relevant paper:

[2101.05667] The Expando-Mono-Duo Design Pattern for Text Ranking with Pretrained Sequence-to-Sequence Models

@ARTICLE{pradeep2021emd,
  title = {The Expando-Mono-Duo Design Pattern for Text Ranking with Pretrained Sequence-to-Sequence Models},
  author = {Ronak Pradeep and Rodrigo Nogueira and Jimmy Lin},
  year = {2021},
  journal = {arXiv:2101.05667},
}

🙏 Acknowledgments

This research is supported in part by the Natural Sciences and Engineering Research Council (NSERC) of Canada.