microsoft / ANCE

A novel embedding training algorithm leveraging ANN search and achieved SOTA retrieval on Trec DL 2019 and OpenQA benchmarks
MIT License
359 stars 49 forks source link

Can't reproduce the performance of warmup(60k) #22

Open xsc1234 opened 1 year ago

xsc1234 commented 1 year ago

Hello, I used run_train_warmup.sh to train the warmup model and found that the performance of my model can not achieve the effect of your released checkpoint (pretrained BM25 warmup checkpoint MRR@10 is 0.311), even if I train it to 300k steps (MRR@10 is 0.2979). All training hyperparameters are as described in run_train_warmup.sh, how can I handle this?