This is the official PyTorch implementation of paper CLUSTERLLM: Large Language Models as a Guide for Text Clustering (EMNLP2023).

pip install -r requirements.txt

Download zip file here and unzip.

cd perspective/2_finetune
bash scripts/get_embedding.sh

The embeddings are produced in each folder of datasets. It will also save the clustering measures. Details instructions see bash script. E5 embeddings are produced with scripts/get_embedding_e5.sh.

cd perspective/1_predict_triplet
bash scripts/triplet_sampling.sh

Sampled triplets will be produced in perspective/1_predict_triplet/sampled_triplet_results. Details instructions see bash script.

First replace the openai keys in perspective/1_predict_triplet/scripts/predict_triplet.sh.

cd perspective/1_predict_triplet
bash scripts/predict_triplet.sh

Predicted triplets will be in perspective/1_predict_triplet/predicted_triplet_results. Details instructions see bash script.

This step only converts the format.

cd perspective/2_finetune
bash scripts/convert_triplet.sh
bash scripts/convert_triplet_self.sh

Converted triplets will be in perspective/2_finetune/converted_triplet_results. Details instructions see bash script.

cd perspective/2_finetune
bash scripts/finetune.sh

Finetuned model will be in perspective/2_finetune/checkpoints. Details instructions see bash script.

cd perspective/2_finetune
bash scripts/get_embedding.sh

This time, switch to checkpoints. Clustering measures will be saved into checkpoint folder.

cd granularity
bash scripts/sample_pairs.sh

Sampled pairs will be saved in sampled_pair_results.

4 pairs will be sampled as in-context examples.

cd granularity
bash scripts/sample_pairs_for_prompt.sh

First replace the openai keys in granularity/scripts/predict_pairs.sh.

cd granularity
bash scripts/predict_pairs.sh

Predicted pairs will be in granularity/predicted_pair_results. Also specify prompt_file to sampled the prompt.

cd granularity
bash scripts/predict_num_clusters.sh

Details instructions see bash script.

I have reduced the run for the whole pipeline into 2 bash scrips.

First, run create_output_folders.sh once to create a list of emptied folder to store ouput of each step in the pipeline:

bash create_output_folders.sh

Then run the pipeline,

bash run_file.sh -i 3 -e 4

For the first run, please run the pipeline with a small number of epoch and iteration to test the pipline. The default iteration (-i) is 6 and epoch (-e) is 30.

IMPORTANT

Please run these 2 files in ClusterLLM or else the path's references in the bash files gonna be incorrect. For example:

cd ClusterLLM
bash run_file.sh

@misc{zhang2023clusterllm,
      title={ClusterLLM: Large Language Models as a Guide for Text Clustering}, 
      author={Yuwei Zhang and Zihan Wang and Jingbo Shang},
      year={2023},
      eprint={2305.14871},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Some of the code was adapted from:

• https://github.com/xlang-ai/instructor-embedding

Yuwei Zhang [email protected]