This is the official code for paper RankingGPT: Empowering Large Language Models in Text Ranking with Progressive Enhancement.

transformers==4.28.1
datasets
pyserini
torch==1.13.1

• ./datasets/text_pairs.json: Weakly supervised text pairs

• ./datasets/msmarco.json: Supervised fine-tuning data

• ./rankdata/trec19: Top-1000 query-document pairs recalled by BM25

bash pretrain.sh bigscience/bloom-560m bloom-560m BloomBlock

bash sft.sh ./outputs_pretrain_bloom-560m bloom-560m 16 BloomBlock

bash eval.sh ./outputs_sft_bloom-560m trec19 bloom-560m

Ranking results (NDCG@10) of the top-1000 candidate documents recalled by BM25.